[doc] update docs (#2692)

.flake8

@@ -1,7 +0,0 @@
-[flake8]
-ignore = E203, E402, E501, E731, E741, W503, W605, E722, E231, W604, E702, E226, E221, E713, E271
-max-line-length = 119
-
-# E402: module level import not at top of file
-per-file-ignores =
-    __init__.py:F401,F403,E402

.github/workflows/Codestyle-Check.yml

@@ -1,51 +0,0 @@
-name: Codestyle-Check
-
-on:
-  pull_request:
-    branches:
-      - develop
-      - 'release/*'
-      - 'feature/*'
-
-jobs:
-  pre-commit:
-    name: Pre Commit
-    if: ${{ github.repository_owner == 'PaddlePaddle' }}
-    runs-on: ubuntu-latest
-    env:
-      PR_ID: ${{ github.event.pull_request.number }}
-      BRANCH: ${{ github.event.pull_request.base.ref }}
-
-    steps:
-      - name: Cleanup
-        run: |
-          rm -rf * .[^.]*
-
-      - name: Checkout base repo
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event.pull_request.base.ref }}
-          fetch-depth: 1000
-
-      - name: Merge PR to test branch
-        run: |
-          git fetch origin pull/${PR_ID}/merge
-          git checkout -b test FETCH_HEAD
-
-      - name: Setup python3.10
-        uses: actions/setup-python@v5
-        with:
-          python-version: '3.10'
-          cache: 'pip'
-
-      - name: Install dependencies
-        run: |
-          pip install pre-commit==4.2.0 cpplint==1.6.0 clang-format==13.0.0
-
-      - name: Check pre-commit
-        env:
-          SKIP_CLANG_TIDY_CHECK: "ON"
-        run: |
-          set +e
-          bash -x tools/codestyle/pre_commit.sh;EXCODE=$?
-          exit $EXCODE

.github/workflows/_accuracy_test.yml

@@ -1,186 +0,0 @@
-name: Accuracy Test
-description: "Run Accuracy Tests"
-
-on:
-  workflow_call:
-    inputs:
-      DOCKER_IMAGE:
-        description: "Build Images"
-        required: true
-        type: string
-        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:cuda126-py310"
-      FASTDEPLOY_ARCHIVE_URL:
-        description: "URL of the compressed FastDeploy code archive."
-        required: true
-        type: string
-      FASTDEPLOY_WHEEL_URL:
-        description: "URL of the FastDeploy Wheel."
-        required: true
-        type: string
-      CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-      MODEL_CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-
-jobs:
-  accuracy_tests:
-    runs-on: [self-hosted, GPU-h20-1Cards]
-    timeout-minutes: 60
-    steps:
-      - name: Code Prepare
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
-        run: |
-            set -x
-            REPO="https://github.com/${{ github.repository }}.git"
-            FULL_REPO="${{ github.repository }}"
-            REPO_NAME="${FULL_REPO##*/}"
-            BASE_BRANCH="${{ github.base_ref }}"
-
-            # Clean the repository directory before starting
-            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
-            -e "REPO_NAME=${REPO_NAME}" \
-            ${docker_image} /bin/bash -c '
-              if [ -d ${REPO_NAME} ]; then
-                echo "Directory ${REPO_NAME} exists, removing it..."
-                rm -rf ${REPO_NAME}*
-              fi
-            '
-
-            wget -q ${fd_archive_url}
-            tar -xf FastDeploy.tar.gz
-            rm -rf FastDeploy.tar.gz
-            cd FastDeploy
-            git config --global user.name "FastDeployCI"
-            git config --global user.email "fastdeploy_ci@example.com"
-            git log -n 3 --oneline
-
-      - name: Run FastDeploy Base Tests
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          fastdeploy_wheel_url: ${{ inputs.FASTDEPLOY_WHEEL_URL }}
-          CACHE_DIR: ${{ inputs.CACHE_DIR }}
-          MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
-        run: |
-          runner_name="${{ runner.name }}"
-          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
-          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
-          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
-
-          FLASK_PORT=$((42068 + DEVICE_PORT * 100))
-          FD_API_PORT=$((42088 + DEVICE_PORT * 100))
-          FD_ENGINE_QUEUE_PORT=$((42058 + DEVICE_PORT * 100))
-          FD_METRICS_PORT=$((42078 + DEVICE_PORT * 100))
-          FD_CACHE_QUEUE_PORT=$((42098 + DEVICE_PORT * 100))
-          echo "Test ENV Parameter:"
-          echo "========================================================="
-          echo "FLASK_PORT=${FLASK_PORT}"
-          echo "FD_API_PORT=${FD_API_PORT}"
-          echo "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}"
-          echo "FD_METRICS_PORT=${FD_METRICS_PORT}"
-          echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
-          echo "DEVICES=${DEVICES}"
-          echo "========================================================="
-
-          CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
-          echo "CACHE_DIR is set to ${CACHE_DIR}"
-          if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
-            touch "${CACHE_DIR}/gitconfig"
-          fi
-          if [ ! -d "${MODEL_CACHE_DIR}" ]; then
-            echo "Error: MODEL_CACHE_DIR '${MODEL_CACHE_DIR}' does not exist."
-            exit 1
-          fi
-
-          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT $FD_CACHE_QUEUE_PORT)
-          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
-          echo "==== LOG_FILE is ${LOG_FILE} ===="
-
-          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
-
-          for port in "${PORTS[@]}"; do
-              PIDS=$(lsof -t -i :$port || true)
-              if [ -n "$PIDS" ]; then
-                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
-                  echo "$PIDS" | xargs -r kill -9
-                  echo "Port $port cleared" | tee -a $LOG_FILE
-              else
-                  echo "Port $port is free" | tee -a $LOG_FILE
-              fi
-          done
-
-          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
-
-          echo "========================================================="
-          echo "Ensuring no stale container named ${runner_name} ..."
-          if [ "$(docker ps -a -q -f name=${runner_name})" ]; then
-            echo "Removing stale container: ${runner_name}"
-            docker rm -f ${runner_name} || true
-          fi
-
-          docker run --rm --ipc=host --pid=host --net=host \
-          --name ${runner_name} \
-          -v $(pwd):/workspace \
-          -w /workspace \
-          -e fastdeploy_wheel_url=${fastdeploy_wheel_url} \
-          -e "FD_API_PORT=${FD_API_PORT}" \
-          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
-          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
-          -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
-          -e "FLASK_PORT=${FLASK_PORT}" \
-          -v "${MODEL_CACHE_DIR}:/MODELDATA" \
-          -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
-          -v "${CACHE_DIR}/.cache:/root/.cache" \
-          -v "${CACHE_DIR}/ConfigDir:/root/.config" \
-          -e TZ="Asia/Shanghai" \
-          --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
-          python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
-
-          pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
-
-          python -m pip install ${fastdeploy_wheel_url}
-          python -m pip install pytest
-
-          wget https://paddle-qa.bj.bcebos.com/zhengtianyu/tools/llm-deploy-linux-amd64
-          chmod +x ./llm-deploy-linux-amd64
-          ./llm-deploy-linux-amd64 -python python3.10 \
-          -model_name ERNIE-4.5-0.3B-Paddle \
-          -model_path /MODELDATA \
-          --skip install
-
-          git config --global --add safe.directory /workspace/FastDeploy
-          cd FastDeploy
-          pushd tests/ce/deploy
-          python3.10 deploy.py > dd.log 2>&1 &
-          sleep 3
-          curl -X POST http://0.0.0.0:${FLASK_PORT}/start \
-            -H "Content-Type: application/json" \
-            -d "{\"--model\": \"/MODELDATA/ERNIE-4.5-0.3B-Paddle\"}"
-
-          curl -X POST http://localhost:${FLASK_PORT}/wait_for_infer?timeout=90
-          popd
-
-          pushd tests/ce/accuracy_cases
-          export URL=http://localhost:${FD_API_PORT}/v1/chat/completions
-          export TEMPLATE=TOKEN_LOGPROB
-          export MODEL_SIZE=0.3B
-          TEST_EXIT_CODE=0
-          python gsm8k.py || TEST_EXIT_CODE=1
-          popd
-          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}" >> /workspace/FastDeploy/exit_code.env
-          '
-          if [ -f ./FastDeploy/exit_code.env ]; then
-            source ./FastDeploy/exit_code.env
-            cat ./FastDeploy/exit_code.env >> $GITHUB_ENV
-          fi
-          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}"
-          exit ${TEST_EXIT_CODE}

.github/workflows/_base_test.yml

@@ -1,229 +0,0 @@
-name: Base Test
-description: "Run Base Tests"
-
-on:
-  workflow_call:
-    inputs:
-      DOCKER_IMAGE:
-        description: "Build Images"
-        required: true
-        type: string
-        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:cuda126-py310"
-      FASTDEPLOY_ARCHIVE_URL:
-        description: "URL of the compressed FastDeploy code archive."
-        required: true
-        type: string
-      FASTDEPLOY_WHEEL_URL:
-        description: "URL of the FastDeploy Wheel."
-        required: true
-        type: string
-      CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-      MODEL_CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-
-jobs:
-  base_tests:
-    runs-on: [self-hosted, GPU-h20-1Cards]
-    timeout-minutes: 60
-    steps:
-      - name: Code Prepare
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
-        run: |
-            set -x
-            REPO="https://github.com/${{ github.repository }}.git"
-            FULL_REPO="${{ github.repository }}"
-            REPO_NAME="${FULL_REPO##*/}"
-            BASE_BRANCH="${{ github.base_ref }}"
-
-            # Clean the repository directory before starting
-            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
-            -e "REPO_NAME=${REPO_NAME}" \
-            ${docker_image} /bin/bash -c '
-              if [ -d ${REPO_NAME} ]; then
-                echo "Directory ${REPO_NAME} exists, removing it..."
-                rm -rf ${REPO_NAME}*
-              fi
-            '
-
-            wget -q ${fd_archive_url}
-            tar -xf FastDeploy.tar.gz
-            rm -rf FastDeploy.tar.gz
-            cd FastDeploy
-            git config --global user.name "FastDeployCI"
-            git config --global user.email "fastdeploy_ci@example.com"
-            git log -n 3 --oneline
-
-      - name: Run FastDeploy Base Tests
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          fastdeploy_wheel_url: ${{ inputs.FASTDEPLOY_WHEEL_URL }}
-          CACHE_DIR: ${{ inputs.CACHE_DIR }}
-          MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
-        run: |
-          runner_name="${{ runner.name }}"
-          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
-          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
-          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
-
-          FLASK_PORT=$((42068 + DEVICE_PORT * 100))
-          FD_API_PORT=$((42088 + DEVICE_PORT * 100))
-          FD_ENGINE_QUEUE_PORT=$((42058 + DEVICE_PORT * 100))
-          FD_METRICS_PORT=$((42078 + DEVICE_PORT * 100))
-          FD_CACHE_QUEUE_PORT=$((42098 + DEVICE_PORT * 100))
-          echo "Test ENV Parameter:"
-          echo "========================================================="
-          echo "FLASK_PORT=${FLASK_PORT}"
-          echo "FD_API_PORT=${FD_API_PORT}"
-          echo "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}"
-          echo "FD_METRICS_PORT=${FD_METRICS_PORT}"
-          echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
-          echo "DEVICES=${DEVICES}"
-          echo "========================================================="
-
-          CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
-          echo "CACHE_DIR is set to ${CACHE_DIR}"
-          if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
-            touch "${CACHE_DIR}/gitconfig"
-          fi
-          if [ ! -d "${MODEL_CACHE_DIR}" ]; then
-            echo "Error: MODEL_CACHE_DIR '${MODEL_CACHE_DIR}' does not exist."
-            exit 1
-          fi
-
-          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT $FD_CACHE_QUEUE_PORT)
-          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
-          echo "==== LOG_FILE is ${LOG_FILE} ===="
-
-          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
-
-          for port in "${PORTS[@]}"; do
-              PIDS=$(lsof -t -i :$port || true)
-              if [ -n "$PIDS" ]; then
-                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
-                  echo "$PIDS" | xargs -r kill -9
-                  echo "Port $port cleared" | tee -a $LOG_FILE
-              else
-                  echo "Port $port is free" | tee -a $LOG_FILE
-              fi
-          done
-
-          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
-
-          echo "========================================================="
-          echo "Ensuring no stale container named ${runner_name} ..."
-          if [ "$(docker ps -a -q -f name=${runner_name})" ]; then
-            echo "Removing stale container: ${runner_name}"
-            docker rm -f ${runner_name} || true
-          fi
-
-          docker run --rm --ipc=host --pid=host --net=host \
-          --name ${runner_name} \
-          -v $(pwd):/workspace \
-          -w /workspace \
-          -e fastdeploy_wheel_url=${fastdeploy_wheel_url} \
-          -e "FD_API_PORT=${FD_API_PORT}" \
-          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
-          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
-          -e "FLASK_PORT=${FLASK_PORT}" \
-          -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
-          -v "${MODEL_CACHE_DIR}:/MODELDATA" \
-          -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
-          -v "${CACHE_DIR}/.cache:/root/.cache" \
-          -v "${CACHE_DIR}/ConfigDir:/root/.config" \
-          -e TZ="Asia/Shanghai" \
-          --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
-          python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
-
-          pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
-
-          python -m pip install ${fastdeploy_wheel_url}
-          python -m pip install pytest
-
-          wget https://paddle-qa.bj.bcebos.com/zhengtianyu/tools/llm-deploy-linux-amd64
-          chmod +x ./llm-deploy-linux-amd64
-          ./llm-deploy-linux-amd64 -python python3.10 \
-          -model_name ERNIE-4.5-0.3B-Paddle \
-          -model_path /MODELDATA \
-          --skip install
-
-          git config --global --add safe.directory /workspace/FastDeploy
-          cd FastDeploy
-          pushd tests/ce/deploy
-          python3.10 deploy.py > dd.log 2>&1 &
-          sleep 3
-          curl -X POST http://0.0.0.0:${FLASK_PORT}/start \
-            -H "Content-Type: application/json" \
-            -d "{\"--model\": \"/MODELDATA/ERNIE-4.5-0.3B-Paddle\"}"
-
-          check_service() {
-            local timeout=${1:-90}
-            local url="http://localhost:${FLASK_PORT}/wait_for_infer?timeout=${timeout}"
-            local resp
-
-            resp=$(curl -s -X POST "$url")
-
-            if echo "$resp" | grep -q "服务启动超时"; then
-              exit 8
-            fi
-          }
-
-          check_service 90
-          popd
-
-          pushd tests/ce/server
-          export URL=http://localhost:${FD_API_PORT}/v1/chat/completions
-          export TEMPLATE=TOKEN_LOGPROB
-          TEST_EXIT_CODE=0
-          python -m pytest -sv test_base_chat.py test_compare_top_logprobs.py test_logprobs.py test_params_boundary.py test_seed_usage.py test_stream.py test_evil_cases.py test_completions.py test_return_token_ids.py || TEST_EXIT_CODE=1
-          curl -X POST http://0.0.0.0:${FLASK_PORT}/switch \
-            -H "Content-Type: application/json" \
-            -d "{\"--model\": \"/MODELDATA/ERNIE-4.5-0.3B-Paddle\", \"--early-stop-config\": \"{\\\"enable_early_stop\\\":true, \\\"window_size\\\":6, \\\"threshold\\\":0.93}\"}"
-          check_service 90
-          python -m pytest -sv test_repetition_early_stop.py || TEST_EXIT_CODE=1
-
-          curl -X POST http://0.0.0.0:${FLASK_PORT}/switch \
-            -H "Content-Type: application/json" \
-            -d "{ \"--model\": \"/MODELDATA/ERNIE-4.5-0.3B-Paddle\", \"--max-concurrency\": 5, \"--max-waiting-time\": 1 }"
-          check_service 90
-          python -m pytest -sv test_max_concurrency.py || TEST_EXIT_CODE=1
-
-          curl -X POST http://0.0.0.0:${FLASK_PORT}/switch \
-            -H "Content-Type: application/json" \
-            -d "{ \"--model\": \"/MODELDATA/ERNIE-4.5-0.3B-Paddle\", \"--max-concurrency\": 5000, \"--max-waiting-time\": 1 }"
-          check_service 90
-          python -m pytest -sv test_max_waiting_time.py || TEST_EXIT_CODE=1
-
-          curl -X POST http://0.0.0.0:${FLASK_PORT}/switch \
-            -H "Content-Type: application/json" \
-            -d "{\"--model\": \"/MODELDATA/ernie-4_5-21b-a3b-bf16-paddle\", \"--config\": \"21b_mtp.yaml\", \"--enable-logprob\": \"False\"}"
-          check_service 180
-          export TEMPLATE=TOKEN_NORMAL
-          python -m pytest -sv test_seed_usage.py -k "not test_seed_stream" || TEST_EXIT_CODE=1
-
-          curl -X POST http://0.0.0.0:${FLASK_PORT}/switch \
-            -H "Content-Type: application/json" \
-            -d "{\"--model\": \"/MODELDATA/ernie-4_5-21b-a3b-bf16-paddle\", \"--config\": \"21b_sot.yaml\", \"--enable-logprob\": \"False\"}"
-          check_service 360
-          export TEMPLATE=TOKEN_NORMAL
-          python -m pytest -sv test_seed_usage.py -k "not test_seed_stream" || TEST_EXIT_CODE=1
-
-          popd
-          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}" >> /workspace/FastDeploy/exit_code.env
-          '
-          if [ -f ./FastDeploy/exit_code.env ]; then
-            source ./FastDeploy/exit_code.env
-            cat ./FastDeploy/exit_code.env >> $GITHUB_ENV
-          fi
-          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}"
-          exit ${TEST_EXIT_CODE}

.github/workflows/_build_linux.yml

@@ -1,199 +0,0 @@
-name: FastDeploy Linux GPU Build Task
-description: "FastDeploy packages build and upload"
-
-on:
-  workflow_call:
-    inputs:
-      DOCKER_IMAGE:
-        description: "Build Images"
-        required: true
-        type: string
-        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:cuda126-py310"
-      FASTDEPLOY_ARCHIVE_URL:
-        description: "URL of the compressed FastDeploy code archive."
-        required: true
-        type: string
-      COMPILE_ARCH:
-        description: "Build GPU Archs"
-        required: true
-        type: string
-        default: "80,90"
-      WITH_NIGHTLY_BUILD:
-        description: "Enable nightly build mode (e.g. add date suffix to version)"
-        required: false
-        type: string
-        default: "OFF"
-      FD_VERSION:
-        description: "FastDeploy Package Version"
-        required: false
-        type: string
-        default: ""
-      PADDLEVERSION:
-        description: "Paddle Version Build Use"
-        required: false
-        type: string
-        default: ""
-      PADDLE_WHL_URL:
-        description: "Paddle Wheel Package URL"
-        required: false
-        type: string
-        default: ""
-      UPLOAD:
-        description: "Upload Package"
-        required: false
-        type: string
-        default: "ON"
-      CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-    outputs:
-      wheel_path:
-        description: "Output path of the generated wheel"
-        value: ${{ jobs.fd-build.outputs.wheel_path }}
-jobs:
-  fd-build:
-    runs-on: [self-hosted, GPU-Build]
-    timeout-minutes: 240
-    outputs:
-      wheel_path: ${{ steps.set_output.outputs.wheel_path }}
-    steps:
-      - name: Code Prepare
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
-          IS_PR: ${{ github.event_name == 'pull_request' }}
-        run: |
-            set -x
-            REPO="https://github.com/${{ github.repository }}.git"
-            FULL_REPO="${{ github.repository }}"
-            REPO_NAME="${FULL_REPO##*/}"
-            BASE_BRANCH="${{ github.base_ref }}"
-
-            # Clean the repository directory before starting
-            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
-            -e "REPO_NAME=${REPO_NAME}" \
-            ${docker_image} /bin/bash -c '
-              if [ -d ${REPO_NAME} ]; then
-                echo "Directory ${REPO_NAME} exists, removing it..."
-                rm -rf ${REPO_NAME}*
-              fi
-            '
-
-            wget -q ${fd_archive_url}
-            tar -xf FastDeploy.tar.gz
-            rm -rf FastDeploy.tar.gz
-            cd FastDeploy
-            git config --global user.name "FastDeployCI"
-            git config --global user.email "fastdeploy_ci@example.com"
-            git log -n 3 --oneline
-      - name: FastDeploy Build
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          compile_arch: ${{ inputs.COMPILE_ARCH }}
-          fd_version: ${{ inputs.FD_VERSION }}
-          CACHE_DIR: ${{ inputs.CACHE_DIR }}
-          BRANCH_REF: ${{ github.ref_name }}
-          PADDLEVERSION: ${{ inputs.PADDLEVERSION }}
-          PADDLE_WHL_URL: ${{ inputs.PADDLE_WHL_URL }}
-          WITH_NIGHTLY_BUILD: ${{ inputs.WITH_NIGHTLY_BUILD }}
-        run: |
-            set -x
-            runner_name="${{ runner.name }}"
-            CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
-            gpu_id=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
-
-            CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
-            echo "CACHE_DIR is set to ${CACHE_DIR}"
-            if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
-              touch "${CACHE_DIR}/gitconfig"
-            fi
-            PARENT_DIR=$(dirname "$WORKSPACE")
-            echo "PARENT_DIR:$PARENT_DIR"
-            docker run --rm --net=host \
-            --cap-add=SYS_PTRACE --privileged --shm-size=64G \
-            -v $(pwd):/workspace -w /workspace \
-            -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
-            -v "${CACHE_DIR}/.cache:/root/.cache" \
-            -v "${CACHE_DIR}/.ccache:/root/.ccache" \
-            -v "${CACHE_DIR}/ConfigDir:/root/.config" \
-            -e TZ="Asia/Shanghai" \
-            -e "COMPILE_ARCH=${compile_arch}" \
-            -e "FD_VERSION=${fd_version}" \
-            -e "WITH_NIGHTLY_BUILD=${WITH_NIGHTLY_BUILD}" \
-            -e "PADDLEVERSION=${PADDLEVERSION}" \
-            -e "PADDLE_WHL_URL=${PADDLE_WHL_URL}" \
-            -e "BRANCH_REF=${BRANCH_REF}" \
-            --gpus "\"device=${gpu_id}\"" ${docker_image} /bin/bash -c '
-            if [[ -n "${FD_VERSION}" ]]; then
-              export FASTDEPLOY_VERSION=${FD_VERSION}
-              echo "Custom FastDeploy version: ${FASTDEPLOY_VERSION}"
-            fi
-
-            git config --global --add safe.directory /workspace/FastDeploy
-            cd FastDeploy
-            if [[ "${WITH_NIGHTLY_BUILD}" == "ON" ]];then
-              GIT_COMMIT_TIME=$(git --no-pager show -s --format=%ci HEAD)
-              DATE_ONLY=$(echo $GIT_COMMIT_TIME | sed "s/ .*//;s/-//g")
-              echo "Git Commit Time: $GIT_COMMIT_TIME"
-              echo "Date Only: $DATE_ONLY"
-              export FASTDEPLOY_VERSION="${FASTDEPLOY_VERSION}.dev${DATE_ONLY}"
-            fi
-            # 针对不同分支和tag使用不同的PaddlePaddle安装包
-            if [[ "${PADDLE_WHL_URL}" != "" ]];then
-              python -m pip install ${PADDLE_WHL_URL}
-            elif [[ "${PADDLEVERSION}" != "" ]];then
-              python -m pip install paddlepaddle-gpu==${PADDLEVERSION} -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
-            else
-              python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
-            fi
-
-            pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
-
-            python -m pip install --upgrade pip
-            python -m pip install -r requirements.txt
-            python -m pip install wheel
-            # 编译RDMA
-            export ENABLE_FD_RDMA=1
-            bash build.sh 1 python false [${COMPILE_ARCH}]
-            ls ./dist/*.whl
-            '
-      - name: Package Upload
-        id: set_output
-        env:
-          compile_arch: ${{ inputs.COMPILE_ARCH }}
-        run: |
-            set -x
-            if [[ "${{ github.event_name }}" == "pull_request" ]];then
-              commit_id=${{ github.event.pull_request.head.sha }}
-              pr_num=${{ github.event.pull_request.number }}
-              target_path=paddle-github-action/PR/FastDeploy/${pr_num}/${commit_id}/SM${compile_arch//,/_}
-            elif [[ "${{ github.ref_type }}" == "tag" ]]; then
-              commit_id=${{ github.sha }}
-              tag_name=${{ github.ref_name }}
-              target_path=paddle-github-action/TAG/FastDeploy/${tag_name}/${commit_id}/SM${compile_arch//,/_}
-            else
-              commit_id=${{ github.sha }}
-              branch_name=${{ github.ref_name }}
-              target_path=paddle-github-action/BRANCH/FastDeploy/${branch_name}/${commit_id}/SM${compile_arch//,/_}
-            fi
-            wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
-            push_file=$(realpath bos_tools.py)
-            python --version
-            python -m pip install bce-python-sdk==0.9.29
-            cd FastDeploy/dist/
-            matches=($(ls fastdeploy*.whl))
-            if [ ${#matches[@]} -ne 1 ]; then
-              echo "Error: Found ${#matches[@]} matching files, expected exactly 1"
-              exit 1
-            fi
-            fd_wheel_name=${matches[0]}
-            echo "Found: $fd_wheel_name"
-            tree -L 3
-            python ${push_file} fastdeploy*.whl ${target_path}
-            target_path_stripped="${target_path#paddle-github-action/}"
-            WHEEL_PATH=https://paddle-github-action.bj.bcebos.com/${target_path_stripped}/${fd_wheel_name}
-            echo "wheel_path=${WHEEL_PATH}" >> $GITHUB_OUTPUT

.github/workflows/_clone_linux.yml

@@ -1,78 +0,0 @@
-name: FastDeploy Code Clone
-description: "FastDeploy clone and upload"
-
-on:
-  workflow_call:
-    inputs:
-        bos_dir:
-          type: string
-          required: false
-          default: 'FastDeploy'
-    outputs:
-      repo_archive_url:
-        description: "Compressed source code archive."
-        value: ${{ jobs.code-clone.outputs.repo_archive_url }}
-jobs:
-  code-clone:
-    runs-on:
-      group: HK-Clone
-    outputs:
-      repo_archive_url: ${{ steps.set_output.outputs.repo_archive_url }}
-    steps:
-      - name: Clone FastDeploy
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event_name == 'pull_request'
-                && github.event.pull_request.base.ref
-                || github.ref_name }}
-          submodules: 'recursive'
-          fetch-depth: 1000
-
-      - name: Merge PR (if needed)
-        if: ${{ github.event_name == 'pull_request' }}
-        run: |
-          git config --global user.name "FastDeployCI"
-          git config --global user.email "fastdeploy_ci@example.com"
-          echo "Fetching and merging PR..."
-          git fetch origin pull/${{ github.event.pull_request.number }}/head:pr/${{ github.event.pull_request.number }}
-          git merge --no-ff pr/${{ github.event.pull_request.number }}
-          echo "PR Branch log "
-          git log --oneline -n 5 pr/${{ github.event.pull_request.number }}
-      - uses: actions/setup-python@v5
-        with:
-          python-version: '3.10'
-      - name: Code Info Show and Upload
-        id: set_output
-        env:
-          AK: paddle
-          SK: paddle
-        run: |
-          git config --unset http.https://github.com/.extraheader
-          git submodule foreach --recursive sh -c "git config --local --unset-all 'http.https://github.com/.extraheader'"
-          git submodule foreach --recursive sh -c "git config remote.origin.fetch '+refs/heads/*:refs/remotes/origin/*'"
-          echo "Current HEAD Log:"
-          git log --oneline -n 5
-          ls
-          cd ..
-          tar -zcf FastDeploy.tar.gz FastDeploy
-          if [[ "${{ github.event_name }}" == "pull_request" ]];then
-            commit_id=${{ github.event.pull_request.head.sha }}
-            pr_num=${{ github.event.pull_request.number }}
-            target_path=paddle-github-action/PR/FastDeploy/${pr_num}/${commit_id}
-          elif [[ "${{ github.ref_type }}" == "tag" ]]; then
-            commit_id=${{ github.sha }}
-            tag_name=${{ github.ref_name }}
-            target_path=paddle-github-action/TAG/FastDeploy/${tag_name}/${commit_id}
-          else
-            commit_id=${{ github.sha }}
-            branch_name=${{ github.ref_name }}
-            target_path=paddle-github-action/BRANCH/FastDeploy/${branch_name}/${commit_id}
-          fi
-          wget -O bos_tools.py -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
-          push_file=$(realpath bos_tools.py)
-          python -m pip install bce-python-sdk==0.9.29
-          ls
-          python ${push_file} FastDeploy.tar.gz ${target_path}
-          target_path_stripped="${target_path#paddle-github-action/}"
-          REPO_ARCHIVE_URL=https://paddle-github-action.bj.bcebos.com/${target_path_stripped}/FastDeploy.tar.gz
-          echo "repo_archive_url=${REPO_ARCHIVE_URL}" >> $GITHUB_OUTPUT

.github/workflows/_logprob_test_linux.yml

@@ -1,184 +0,0 @@
-name: Run FastDeploy LogProb Tests
-description: "Run FastDeploy LogProb Tests"
-
-on:
-  workflow_call:
-    inputs:
-      DOCKER_IMAGE:
-        description: "Build Images"
-        required: true
-        type: string
-        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:cuda126-py310"
-      PADDLETEST_ARCHIVE_URL:
-        description: "URL of the compressed FastDeploy code archive."
-        required: true
-        type: string
-        default: "https://xly-devops.bj.bcebos.com/PaddleTest/PaddleTest.tar.gz"
-      FASTDEPLOY_WHEEL_URL:
-        description: "URL of the FastDeploy Wheel."
-        required: true
-        type: string
-      CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-      MODEL_CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-
-jobs:
-  run_tests_logprob:
-    runs-on: [self-hosted, GPU-h20-1Cards]
-    steps:
-      - name: Code Prepare
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          paddletest_archive_url: ${{ inputs.PADDLETEST_ARCHIVE_URL }}
-        run: |
-            # Clean the repository directory before starting
-            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
-            -e "REPO_NAME=${REPO_NAME}" \
-            -e "BASE_BRANCH=${BASE_BRANCH}" \
-            ${docker_image} /bin/bash -c '
-            rm -rf /workspace/*
-            '
-            wget -q ${paddletest_archive_url}
-            tar -xf PaddleTest.tar.gz
-            rm -rf PaddleTest.tar.gz
-            cd PaddleTest
-            git config --global user.name "FastDeployCI"
-            git config --global user.email "fastdeploy_ci@example.com"
-            git log -n 3 --oneline
-      - name: logprob test
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          fastdeploy_wheel_url: ${{ inputs.FASTDEPLOY_WHEEL_URL }}
-          CACHE_DIR: ${{ inputs.CACHE_DIR }}
-          MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
-        run: |
-          runner_name="${{ runner.name }}"
-          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
-          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
-          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
-
-          FLASK_PORT=$((42068 + DEVICE_PORT * 100))
-          FD_API_PORT=$((42088 + DEVICE_PORT * 100))
-          FD_ENGINE_QUEUE_PORT=$((42058 + DEVICE_PORT * 100))
-          FD_METRICS_PORT=$((42078 + DEVICE_PORT * 100))
-          FD_CACHE_QUEUE_PORT=$((42098 + DEVICE_PORT * 100))
-          echo "Test ENV Parameter:"
-          echo "========================================================="
-          echo "FLASK_PORT=${FLASK_PORT}"
-          echo "FD_API_PORT=${FD_API_PORT}"
-          echo "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}"
-          echo "FD_METRICS_PORT=${FD_METRICS_PORT}"
-          echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
-          echo "DEVICES=${DEVICES}"
-          echo "========================================================="
-
-          CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
-          echo "CACHE_DIR is set to ${CACHE_DIR}"
-          if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
-            touch "${CACHE_DIR}/gitconfig"
-          fi
-          if [ ! -d "${MODEL_CACHE_DIR}" ]; then
-            echo "Error: MODEL_CACHE_DIR '${MODEL_CACHE_DIR}' does not exist."
-            exit 1
-          fi
-
-          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT $FD_CACHE_QUEUE_PORT)
-          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
-          echo "==== LOG_FILE is ${LOG_FILE} ===="
-
-          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
-
-          for port in "${PORTS[@]}"; do
-              PIDS=$(lsof -t -i :$port || true)
-              if [ -n "$PIDS" ]; then
-                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
-                  echo "$PIDS" | xargs -r kill -9
-                  echo "Port $port cleared" | tee -a $LOG_FILE
-              else
-                  echo "Port $port is free" | tee -a $LOG_FILE
-              fi
-          done
-
-          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
-
-          echo "========================================================="
-          echo "Ensuring no stale container named ${runner_name} ..."
-          if [ "$(docker ps -a -q -f name=${runner_name})" ]; then
-            echo "Removing stale container: ${runner_name}"
-            docker rm -f ${runner_name} || true
-          fi
-
-          docker run --rm --ipc=host --pid=host --net=host \
-          --name ${runner_name} \
-          -v $(pwd):/workspace \
-          -w /workspace \
-          -e fastdeploy_wheel_url=${fastdeploy_wheel_url} \
-          -e "FD_API_PORT=${FD_API_PORT}" \
-          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
-          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
-          -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
-          -e "FLASK_PORT=${FLASK_PORT}" \
-          -v "${MODEL_CACHE_DIR}:/MODELDATA" \
-          -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
-          -v "${CACHE_DIR}/.cache:/root/.cache" \
-          -v "${CACHE_DIR}/ConfigDir:/root/.config" \
-          -e TZ="Asia/Shanghai" \
-          --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
-          python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
-
-          pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
-
-          python -m pip install ${fastdeploy_wheel_url}
-
-          wget https://paddle-qa.bj.bcebos.com/zhengtianyu/tools/llm-deploy-linux-amd64
-          chmod +x ./llm-deploy-linux-amd64
-          ./llm-deploy-linux-amd64 -python python3.10 \
-          -model_name ERNIE-4.5-0.3B-Paddle \
-          -model_path /MODELDATA \
-          --skip install
-
-          cd PaddleTest/framework/ServeTest
-          python3.10 deploy.py > dd.log 2>&1 &
-          sleep 3
-          curl -X POST http://0.0.0.0:${FLASK_PORT}/start \
-              -H "Content-Type: application/json" \
-              -d "{\"--model\": \"/MODELDATA/ERNIE-4.5-0.3B-Paddle\"}"
-
-          curl -X POST http://localhost:${FLASK_PORT}/wait_for_infer?timeout=90
-          curl -s -o /dev/null -w "%{http_code}" -m 2 "http://0.0.0.0:${FD_API_PORT}/health"
-          curl -X POST "http://0.0.0.0:${FD_API_PORT}/v1/chat/completions" \
-            -H "Content-Type: application/json" \
-            -d "{\"messages\": [{\"role\": \"user\", \"content\": \"1+1=?\"}], \"logprobs\": true}"
-          set +e
-          rm -rf ./baseline_output
-          cp -r baseline/ERNIE-4.5-0.3B-Paddle ./baseline_output
-          LOGPROB_EXIT_CODE=0
-          python3.10 lanucher.py --request_template TOKEN_LOGPROB --url http://localhost:${FD_API_PORT}/v1/chat/completions  --case ./cases/demo.yaml  --concurrency 1 --name demo --exe logprob || LOGPROB_EXIT_CODE=$?
-          echo "LOGPROB_EXIT_CODE=${LOGPROB_EXIT_CODE}" > /workspace/exit_code.env
-          curl -X POST http://localhost:${FLASK_PORT}/stop
-          sleep 10s
-          cat *result.log
-          exit 0
-          '
-          if [ $? -ne 0 ];then
-            exit 1
-          fi
-
-          if [ -f exit_code.env ]; then
-            cat exit_code.env >> $GITHUB_ENV
-          fi
-      - name: logprob test result
-        if: ${{ env.LOGPROB_EXIT_CODE != 0 }}
-        shell: bash
-        run: |
-          echo "logprob test failed with exit code ${{ env.LOGPROB_EXIT_CODE }}"
-          exit 8

.github/workflows/_pre_ce_test.yml

@@ -1,148 +0,0 @@
-name: Pre-CE-Test
-
-on:
-  workflow_call:
-    inputs:
-      DOCKER_IMAGE:
-        description: "Build Images"
-        required: true
-        type: string
-        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddle:fastdeploy-ciuse-cuda126"
-      FASTDEPLOY_ARCHIVE_URL:
-        description: "URL of the compressed FastDeploy code archive."
-        required: true
-        type: string
-      FASTDEPLOY_WHEEL_URL:
-        description: "URL of the FastDeploy Wheel."
-        required: true
-        type: string
-      CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-      MODEL_CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-
-jobs:
-  run_ce_cases:
-    runs-on: [self-hosted, PRE_CE_RUN_2Card]
-    timeout-minutes: 60
-    steps:
-      - name: Print current runner name
-        run: |
-          echo "Current runner name: ${{ runner.name }}"
-      - name: Code Prepare
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
-        run: |
-            set -x
-            REPO="https://github.com/${{ github.repository }}.git"
-            FULL_REPO="${{ github.repository }}"
-            REPO_NAME="${FULL_REPO##*/}"
-            BASE_BRANCH="${{ github.base_ref }}"
-
-            # Clean the repository directory before starting
-            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
-            -e "REPO_NAME=${REPO_NAME}" \
-            ${docker_image} /bin/bash -c '
-              if [ -d ${REPO_NAME} ]; then
-                echo "Directory ${REPO_NAME} exists, removing it..."
-                rm -rf ${REPO_NAME}*
-              fi
-            '
-
-            wget -q ${fd_archive_url}
-            tar -xf FastDeploy.tar.gz
-            rm -rf FastDeploy.tar.gz
-            cd FastDeploy
-            git config --global user.name "FastDeployCI"
-            git config --global user.email "fastdeploy_ci@example.com"
-            git log -n 3 --oneline
-
-      - name: Run CI unittest
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          fd_wheel_url: ${{ inputs.FASTDEPLOY_WHEEL_URL }}
-          CACHE_DIR: ${{ inputs.CACHE_DIR }}
-          MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
-        run: |
-          runner_name="${{ runner.name }}"
-          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
-          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
-          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
-
-          FLASK_PORT=$((42068 + DEVICE_PORT * 100))
-          FD_API_PORT=$((42088 + DEVICE_PORT * 100))
-          FD_ENGINE_QUEUE_PORT=$((42058 + DEVICE_PORT * 100))
-          FD_METRICS_PORT=$((42078 + DEVICE_PORT * 100))
-          FD_CACHE_QUEUE_PORT=$((42098 + DEVICE_PORT * 100))
-          echo "Test ENV Parameter:"
-          echo "========================================================="
-          echo "FLASK_PORT=${FLASK_PORT}"
-          echo "FD_API_PORT=${FD_API_PORT}"
-          echo "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}"
-          echo "FD_METRICS_PORT=${FD_METRICS_PORT}"
-          echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
-          echo "DEVICES=${DEVICES}"
-          echo "========================================================="
-
-          CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
-          echo "CACHE_DIR is set to ${CACHE_DIR}"
-          if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
-            touch "${CACHE_DIR}/gitconfig"
-          fi
-
-          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT $FD_CACHE_QUEUE_PORT)
-          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
-          echo "==== LOG_FILE is ${LOG_FILE} ===="
-
-          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
-
-          for port in "${PORTS[@]}"; do
-              PIDS=$(lsof -t -i :$port || true)
-              if [ -n "$PIDS" ]; then
-                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
-                  echo "$PIDS" | xargs -r kill -9
-                  echo "Port $port cleared" | tee -a $LOG_FILE
-              else
-                  echo "Port $port is free" | tee -a $LOG_FILE
-              fi
-          done
-
-          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
-
-          echo "========================================================="
-          echo "Ensuring no stale container named ${runner_name} ..."
-          if [ "$(docker ps -a -q -f name=${runner_name})" ]; then
-            echo "Removing stale container: ${runner_name}"
-            docker rm -f ${runner_name} || true
-          fi
-
-          docker run --rm --net=host \
-          --name ${runner_name} \
-          -v $(pwd):/workspace \
-          -w /workspace \
-          -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
-          -v "${CACHE_DIR}/.cache:/root/.cache" \
-          -v "${CACHE_DIR}/ConfigDir:/root/.config" \
-          -v "${MODEL_CACHE_DIR}:/ModelData:ro" \
-          -e "MODEL_PATH=/ModelData" \
-          -e "FD_API_PORT=${FD_API_PORT}" \
-          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
-          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
-          -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
-          -e "FLASK_PORT=${FLASK_PORT}" \
-          -e "fd_wheel_url=${fd_wheel_url}" \
-          --gpus "\"device=${DEVICES}\"" ${docker_image} /bin/bash -c '
-          git config --global --add safe.directory /workspace/FastDeploy
-          cd FastDeploy
-          python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
-          python -m pip install ${fd_wheel_url}
-          bash scripts/run_pre_ce.sh
-          '

.github/workflows/_stable_test.yml

@@ -1,170 +0,0 @@
-name: Stable Test
-description: "Run Stable Tests"
-
-on:
-  workflow_call:
-    inputs:
-      DOCKER_IMAGE:
-        description: "Build Images"
-        required: true
-        type: string
-        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:cuda126-py310"
-      FASTDEPLOY_ARCHIVE_URL:
-        description: "URL of the compressed FastDeploy code archive."
-        required: true
-        type: string
-      FASTDEPLOY_WHEEL_URL:
-        description: "URL of the FastDeploy Wheel."
-        required: true
-        type: string
-      CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-      MODEL_CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-
-jobs:
-  stable_tests:
-    runs-on: [self-hosted, GPU-h1z1-2Cards]
-    timeout-minutes: 60
-    steps:
-      - name: Code Prepare
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
-        run: |
-            set -x
-            REPO="https://github.com/${{ github.repository }}.git"
-            FULL_REPO="${{ github.repository }}"
-            REPO_NAME="${FULL_REPO##*/}"
-            BASE_BRANCH="${{ github.base_ref }}"
-
-            # Clean the repository directory before starting
-            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
-            -e "REPO_NAME=${REPO_NAME}" \
-            ${docker_image} /bin/bash -c '
-              if [ -d ${REPO_NAME} ]; then
-                echo "Directory ${REPO_NAME} exists, removing it..."
-                rm -rf ${REPO_NAME}*
-              fi
-            '
-
-            wget -q ${fd_archive_url}
-            tar -xf FastDeploy.tar.gz
-            rm -rf FastDeploy.tar.gz
-            cd FastDeploy
-            git config --global user.name "FastDeployCI"
-            git config --global user.email "fastdeploy_ci@example.com"
-            git log -n 3 --oneline
-
-      - name: Run FastDeploy Stable Tests
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          fastdeploy_wheel_url: ${{ inputs.FASTDEPLOY_WHEEL_URL }}
-          CACHE_DIR: ${{ inputs.CACHE_DIR }}
-          MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
-        run: |
-          runner_name="${{ runner.name }}"
-          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
-          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
-          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
-
-          FLASK_PORT=$((42068 + DEVICE_PORT * 100))
-          FD_API_PORT=$((42088 + DEVICE_PORT * 100))
-          FD_ENGINE_QUEUE_PORT=$((42058 + DEVICE_PORT * 100))
-          FD_METRICS_PORT=$((42078 + DEVICE_PORT * 100))
-          FD_CACHE_QUEUE_PORT=$((42038 + DEVICE_PORT * 100))
-          FD_INFERENCE_MSG_QUEUE_ID=$(( 42048 + DEVICE_PORT * 100))
-          echo "Test ENV Parameter:"
-          echo "========================================================="
-          echo "FLASK_PORT=${FLASK_PORT}"
-          echo "FD_API_PORT=${FD_API_PORT}"
-          echo "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}"
-          echo "FD_METRICS_PORT=${FD_METRICS_PORT}"
-          echo "FD_INFERENCE_MSG_QUEUE_ID=${FD_INFERENCE_MSG_QUEUE_ID}"
-          echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
-          echo "DEVICES=${DEVICES}"
-          echo "========================================================="
-
-          CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
-          echo "CACHE_DIR is set to ${CACHE_DIR}"
-          if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
-            touch "${CACHE_DIR}/gitconfig"
-          fi
-          if [ ! -d "${MODEL_CACHE_DIR}" ]; then
-            echo "Error: MODEL_CACHE_DIR '${MODEL_CACHE_DIR}' does not exist."
-            exit 1
-          fi
-
-          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT)
-          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
-          echo "==== LOG_FILE is ${LOG_FILE} ===="
-
-          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
-
-          for port in "${PORTS[@]}"; do
-              PIDS=$(lsof -t -i :$port || true)
-              if [ -n "$PIDS" ]; then
-                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
-                  echo "$PIDS" | xargs -r kill -9
-                  echo "Port $port cleared" | tee -a $LOG_FILE
-              else
-                  echo "Port $port is free" | tee -a $LOG_FILE
-              fi
-          done
-
-          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
-
-          echo "========================================================="
-          echo "Ensuring no stale container named ${runner_name} ..."
-          if [ "$(docker ps -a -q -f name=${runner_name})" ]; then
-            echo "Removing stale container: ${runner_name}"
-            docker rm -f ${runner_name} || true
-          fi
-
-          docker run --rm --ipc=host --pid=host --net=host \
-          --name ${runner_name} \
-          -v $(pwd):/workspace \
-          -w /workspace \
-          -e fastdeploy_wheel_url=${fastdeploy_wheel_url} \
-          -e "FD_API_PORT=${FD_API_PORT}" \
-          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
-          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
-          -e "FLASK_PORT=${FLASK_PORT}" \
-          -e "FD_INFERENCE_MSG_QUEUE_ID=${FD_INFERENCE_MSG_QUEUE_ID}" \
-          -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
-          -v "${MODEL_CACHE_DIR}:/MODELDATA" \
-          -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
-          -v "${CACHE_DIR}/.cache:/root/.cache" \
-          -v "${CACHE_DIR}/ConfigDir:/root/.config" \
-          -e TZ="Asia/Shanghai" \
-          --gpus '"device='"${DEVICES}"'"' ${docker_image} /bin/bash -xc '
-          python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
-
-          pip config set global.index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
-
-          python -m pip install ${fastdeploy_wheel_url}
-          python -m pip install pytest
-
-          git config --global --add safe.directory /workspace/FastDeploy
-          cd FastDeploy
-          TEST_EXIT_CODE=0
-          pushd tests/ce/stable_cases
-          bash launch_model.sh /MODELDATA
-          bash run.sh || TEST_EXIT_CODE=1
-          popd
-          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}" >> /workspace/FastDeploy/exit_code.env
-          '
-          if [ -f ./FastDeploy/exit_code.env ]; then
-            source ./FastDeploy/exit_code.env
-            cat ./FastDeploy/exit_code.env >> $GITHUB_ENV
-          fi
-          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}"
-          exit ${TEST_EXIT_CODE}

.github/workflows/_unit_test_coverage.yml

@@ -1,322 +0,0 @@
-name: Coverage Check
-description: "Run FastDeploy Unit Tests and Coverage"
-
-on:
-  workflow_call:
-    inputs:
-      DOCKER_IMAGE:
-        description: "Build Images"
-        required: true
-        type: string
-        default: "ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:cuda126-py310"
-      FASTDEPLOY_ARCHIVE_URL:
-        description: "URL of the compressed FastDeploy code archive."
-        required: true
-        type: string
-      FASTDEPLOY_WHEEL_URL:
-        description: "URL of the FastDeploy Wheel."
-        required: true
-        type: string
-      CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-      MODEL_CACHE_DIR:
-        description: "Cache Dir Use"
-        required: false
-        type: string
-        default: ""
-    secrets:
-      github-token:
-        required: true
-
-jobs:
-  check_cov_skip:
-    uses: ./.github/workflows/check-bypass.yml
-    secrets:
-      github-token: ${{ secrets.github-token }}
-    with:
-      workflow-name: coverage
-
-  run_tests_with_coverage:
-    runs-on: [self-hosted, GPU-h1z1-2Cards]
-    timeout-minutes: 60
-    needs: check_cov_skip
-    if: needs.check_cov_skip.outputs.can-skip != 'true'
-    outputs:
-      diff_cov_file_url: ${{ steps.cov_upload.outputs.diff_cov_file_url }}
-      unittest_failed_url: ${{ steps.cov_upload.outputs.unittest_failed_url }}
-      diff_cov_result_json_url: ${{ steps.cov_upload.outputs.diff_cov_result_json_url }}
-    steps:
-      - name: Code Prepare
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
-        run: |
-            set -x
-            REPO="https://github.com/${{ github.repository }}.git"
-            FULL_REPO="${{ github.repository }}"
-            REPO_NAME="${FULL_REPO##*/}"
-            BASE_BRANCH="${{ github.base_ref }}"
-
-            # Clean the repository directory before starting
-            docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
-            -e "REPO_NAME=${REPO_NAME}" \
-            ${docker_image} /bin/bash -c '
-              if [ -d ${REPO_NAME} ]; then
-                echo "Directory ${REPO_NAME} exists, removing it..."
-                rm -rf ${REPO_NAME}*
-              fi
-            '
-
-            wget -q ${fd_archive_url}
-            tar -xf FastDeploy.tar.gz
-            rm -rf FastDeploy.tar.gz
-            cd FastDeploy
-            git config --global user.name "FastDeployCI"
-            git config --global user.email "fastdeploy_ci@example.com"
-            git log -n 3 --oneline
-      - name: Run FastDeploy Unit Tests and Coverage
-        shell: bash
-        env:
-          docker_image: ${{ inputs.DOCKER_IMAGE }}
-          fd_wheel_url: ${{ inputs.FASTDEPLOY_WHEEL_URL }}
-          CACHE_DIR: ${{ inputs.CACHE_DIR }}
-          BASE_REF: ${{ github.event.pull_request.base.ref }}
-          MODEL_CACHE_DIR: ${{ inputs.MODEL_CACHE_DIR }}
-          IS_PR: ${{ github.event_name == 'pull_request' }}
-        run: |
-          if [[ "$IS_PR" == "true" ]]; then
-            echo "Running on PR"
-          else
-            echo "Not a PR"
-          fi
-          runner_name="${{ runner.name }}"
-          CARD_ID=$(echo "${runner_name}" | awk -F'-' '{print $NF}')
-          DEVICES=$(echo "$CARD_ID" | fold -w1 | paste -sd,)
-          DEVICE_PORT=$(echo "$DEVICES" | cut -d',' -f1)
-
-          FLASK_PORT=$((42068 + DEVICE_PORT * 100))
-          FD_API_PORT=$((42088 + DEVICE_PORT * 100))
-          FD_ENGINE_QUEUE_PORT=$((42058 + DEVICE_PORT * 100))
-          FD_METRICS_PORT=$((42078 + DEVICE_PORT * 100))
-          FD_CACHE_QUEUE_PORT=$((42098 + DEVICE_PORT * 100))
-          echo "Test ENV Parameter:"
-          echo "========================================================="
-          echo "FLASK_PORT=${FLASK_PORT}"
-          echo "FD_API_PORT=${FD_API_PORT}"
-          echo "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}"
-          echo "FD_METRICS_PORT=${FD_METRICS_PORT}"
-          echo "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}"
-          echo "DEVICES=${DEVICES}"
-          echo "========================================================="
-
-          CACHE_DIR="${CACHE_DIR:-$(dirname "$(dirname "${{ github.workspace }}")")}"
-          echo "CACHE_DIR is set to ${CACHE_DIR}"
-          if [ ! -f "${CACHE_DIR}/gitconfig" ]; then
-            touch "${CACHE_DIR}/gitconfig"
-          fi
-
-          PORTS=($FLASK_PORT $FD_API_PORT $FD_ENGINE_QUEUE_PORT $FD_METRICS_PORT $FD_CACHE_QUEUE_PORT)
-          LOG_FILE="./port_cleanup_$(date +%Y%m%d_%H%M%S).log"
-          echo "==== LOG_FILE is ${LOG_FILE} ===="
-
-          echo "==== PORT CLEAN BEFORE TASK RUN ====" | tee -a $LOG_FILE
-
-          for port in "${PORTS[@]}"; do
-              PIDS=$(lsof -t -i :$port || true)
-              if [ -n "$PIDS" ]; then
-                  echo "Port $port is occupied by PID(s): $PIDS" | tee -a $LOG_FILE
-                  echo "$PIDS" | xargs -r kill -9
-                  echo "Port $port cleared" | tee -a $LOG_FILE
-              else
-                  echo "Port $port is free" | tee -a $LOG_FILE
-              fi
-          done
-
-          echo "==== PORT CLEAN COMPLETE ====" | tee -a $LOG_FILE
-
-          echo "========================================================="
-          echo "Ensuring no stale container named ${runner_name} ..."
-          if [ "$(docker ps -a -q -f name=${runner_name})" ]; then
-            echo "Removing stale container: ${runner_name}"
-            docker rm -f ${runner_name} || true
-          fi
-
-          docker run --rm --net=host \
-          --name ${runner_name} \
-          --cap-add=SYS_PTRACE --shm-size=64G \
-          -v $(pwd):/workspace -w /workspace \
-          -v "${CACHE_DIR}/gitconfig:/etc/gitconfig:ro" \
-          -v "${CACHE_DIR}/.cache:/root/.cache" \
-          -v "${CACHE_DIR}/ConfigDir:/root/.config" \
-          -v "${MODEL_CACHE_DIR}:/ModelData:ro" \
-          -e "MODEL_PATH=/ModelData" \
-          -e "FD_API_PORT=${FD_API_PORT}" \
-          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
-          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
-          -e "FLASK_PORT=${FLASK_PORT}" \
-          -e "FD_CACHE_QUEUE_PORT=${FD_CACHE_QUEUE_PORT}" \
-          -e TZ="Asia/Shanghai" \
-          -e "fd_wheel_url=${fd_wheel_url}" \
-          -e "BASE_REF=${BASE_REF}" \
-          -e "IS_PR=${IS_PR}" \
-          --gpus "\"device=${DEVICES}\"" ${docker_image} /bin/bash -c '
-
-          git config --global --add safe.directory /workspace/FastDeploy
-          cd FastDeploy
-          git diff origin/${BASE_REF}..HEAD --unified=0 > diff.txt
-          python -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu126/
-          pip config set global.extra-index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
-
-          python -m pip install coverage
-          python -m pip install diff-cover
-          python -m pip install pytest-cov
-          python -m pip install jsonschema aistudio_sdk==0.3.5
-          python -m pip install ${fd_wheel_url}
-          rm -rf fastdeploy
-          # coverage subprocess use
-          python -m pip install ${fd_wheel_url} --no-deps --target=/workspace/FastDeploy
-          export PYTHONPATH=/workspace/FastDeploy/
-          if [ -d "tests/plugins" ]; then
-              cd tests/plugins
-              python setup.py install
-              cd ../..
-          else
-              echo "Warning: tests/plugins directory not found, skipping setup.py install"
-          fi
-          export COVERAGE_FILE=/workspace/FastDeploy/coveragedata/.coverage
-          export COVERAGE_RCFILE=/workspace/FastDeploy/scripts/.coveragerc
-          TEST_EXIT_CODE=0
-          bash scripts/coverage_run.sh || TEST_EXIT_CODE=8
-          echo "TEST_EXIT_CODE=${TEST_EXIT_CODE}" >> exit_code.env
-          coverage combine coveragedata/ || echo "No data to combine"
-          coverage report
-          coverage xml -o python_coverage_all.xml
-          COVERAGE_EXIT_CODE=0
-          if [[ "$IS_PR" == "true" ]]; then
-            diff-cover python_coverage_all.xml --diff-file=diff.txt --fail-under=80 --json-report diff_coverage.json || COVERAGE_EXIT_CODE=9
-            python scripts/generate_diff_coverage_xml.py diff.txt python_coverage_all.xml
-          else
-            echo "Not a PR, skipping diff-cover"
-          fi
-          echo "COVERAGE_EXIT_CODE=${COVERAGE_EXIT_CODE}" >> exit_code.env
-          '
-          if [ -f FastDeploy/exit_code.env ]; then
-            cat FastDeploy/exit_code.env >> $GITHUB_ENV
-          fi
-      - name: Upload unit resule and diff coverage to bos
-        id: cov_upload
-        shell: bash
-        run: |
-          cd FastDeploy
-          commit_id=${{ github.event.pull_request.head.sha }}
-          pr_num=${{ github.event.pull_request.number }}
-          target_path=paddle-github-action/PR/FastDeploy/${pr_num}/${commit_id}/SM${compile_arch//,/_}
-          wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py -O bos_tools.py
-          push_file=$(realpath bos_tools.py)
-          python -m pip install bce-python-sdk==0.9.29
-          diff_cov_file="diff_coverage.xml"
-          if [ -f ${diff_cov_file} ];then
-            python ${push_file} ${diff_cov_file} ${target_path}/CoverageData
-            target_path_stripped="${target_path#paddle-github-action/}"
-            DIFF_COV_FILE_URL=https://paddle-github-action.bj.bcebos.com/${target_path_stripped}/CoverageData/${diff_cov_file}
-            echo "diff_cov_file_url=${DIFF_COV_FILE_URL}" >> $GITHUB_OUTPUT
-            echo "diff_cov_file_url=${DIFF_COV_FILE_URL}" >> $GITHUB_ENV
-          fi
-          diff_cov_result_json="diff_coverage.json"
-          if [ -f ${diff_cov_result_json} ];then
-            python ${push_file} ${diff_cov_result_json} ${target_path}/CoverageData
-            target_path_stripped="${target_path#paddle-github-action/}"
-            DIFF_COV_JSON_URL=https://paddle-github-action.bj.bcebos.com/${target_path_stripped}/CoverageData/${diff_cov_result_json}
-            echo "diff_cov_result_json_url=${DIFF_COV_JSON_URL}" >> $GITHUB_OUTPUT
-            echo "diff_cov_result_json_url=${DIFF_COV_JSON_URL}" >> $GITHUB_ENV
-          fi
-          unittest_result="failed_tests.log"
-          if [ -s ${unittest_result} ];then
-            python ${push_file} ${unittest_result} ${target_path}/UnitTestResult
-            target_path_stripped="${target_path#paddle-github-action/}"
-            UNIT_TEST_RESULT_URL=https://paddle-github-action.bj.bcebos.com/${target_path_stripped}/UnitTestResult/${unittest_result}
-            echo "unittest_failed_url=${UNIT_TEST_RESULT_URL}" >> $GITHUB_OUTPUT
-            echo "unittest_failed_url=${UNIT_TEST_RESULT_URL}" >> $GITHUB_ENV
-          fi
-      - name: Check Unit Test Success
-        shell: bash
-        run: |
-          cd FastDeploy
-          if [ "$TEST_EXIT_CODE" -eq 8 ]; then
-            filename=$(basename "$unittest_failed_url")
-            if [ -z "${unittest_failed_url}" ]; then
-              echo "No diff unit failed file URL provided."
-            else
-              rm -rf "${filename}"
-              wget -O ${filename} ${unittest_failed_url} || echo "Download unittest file failed, but continuing..."
-            fi
-            echo "Unit tests failed (exit code 8)"
-            if [ -f "${filename}" ];then
-              echo "Failed test cases:"
-              cat "${filename}"
-            fi
-            exit "$TEST_EXIT_CODE"
-          fi
-          echo "All tests passed"
-
-      - name: Verify Code Coverage Threshold (80%)
-        if: ${{ github.event_name == 'pull_request' }}
-        shell: bash
-        run: |
-          cd FastDeploy
-          if [ "$COVERAGE_EXIT_CODE" -eq 9 ]; then
-            echo "Coverage generation failed (exit code 9)"
-            filename=$(basename "$diff_cov_result_json_url")
-            if [ -z "${diff_cov_result_json_url}" ]; then
-              echo "No diff cov result file URL provided."
-            else
-              rm -rf "${filename}"
-              wget -O ${filename} ${diff_cov_result_json_url} || echo "Download cov json file failed, but continuing..."
-            fi
-            if [ -f "${filename}" ];then
-              echo "Failed test cases:"
-              if command -v jq >/dev/null 2>&1; then
-                  jq . "${filename}"
-              else
-                  cat "${filename}"
-              fi
-            fi
-            exit "$COVERAGE_EXIT_CODE"
-          fi
-          echo "coverage passed"
-          exit 0
-
-  diff_coverage_report:
-    needs: run_tests_with_coverage
-    if: always()
-    runs-on: ubuntu-latest
-    env:
-      fd_archive_url: ${{ inputs.FASTDEPLOY_ARCHIVE_URL }}
-    steps:
-      - name: coverage diff file download
-        shell: bash
-        env:
-          diff_cov_file_url: ${{ needs.run_tests_with_coverage.outputs.diff_cov_file_url }}
-        run: |
-          wget ${fd_archive_url}
-          tar -xf FastDeploy.tar.gz
-          cd FastDeploy
-          if [ -z "${diff_cov_file_url}" ]; then
-            echo "No diff coverage file URL provided."
-            exit 0
-          fi
-          wget "${diff_cov_file_url}" -O ./diff_coverage.xml || echo "Download cov file failed, but continuing..."
-      - name: Upload diff coverage report
-        if: ${{ needs.run_tests_with_coverage.outputs.diff_cov_file_url != null && needs.run_tests_with_coverage.outputs.diff_cov_file_url != '' }}
-        uses: codecov/codecov-action@v5
-        with:
-          files: ./FastDeploy/diff_coverage.xml
-          name: python diff coverage
-          verbose: true
-          disable_search: true
-          commit_parent: false
-          flags: diff

.github/workflows/approve.yml

@@ -1,43 +0,0 @@
-name: Approval
-
-on:
-  pull_request:
-    branches:
-      - develop
-      - 'release/*'
-      - 'feature/*'
-
-env:
-  GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-
-jobs:
-  Approval:
-    name: Approval
-    if: ${{ github.repository_owner == 'PaddlePaddle' }}
-    runs-on: ubuntu-latest
-    env:
-      PR_ID: ${{ github.event.pull_request.number }}
-      BRANCH: ${{ github.event.pull_request.base.ref }}
-    steps:
-      - name: Checkout base repo
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event.pull_request.base.ref }}
-          fetch-depth: 1000
-
-      - name: Merge PR to test branch
-        run: |
-          git fetch origin pull/${PR_ID}/merge
-          git checkout -b test FETCH_HEAD
-          git log -n 3 --oneline
-          git remote add upstream https://github.com/PaddlePaddle/FastDeploy.git
-          git fetch upstream $BRANCH
-
-      - name: Setup python3.10
-        uses: actions/setup-python@v5
-        with:
-          python-version: '3.10'
-
-      - name: Run approval check script
-        run: |
-          bash scripts/check_approval.sh

.github/workflows/ce_job.yml

@@ -1,249 +0,0 @@
-name: CE Compile Job
-
-on:
-  workflow_dispatch:
-  push:
-    branches:
-      - develop
-      - 'release/*'
-      - 'feature/experimental_feature*'
-permissions: read-all
-
-concurrency:
-  group: ${{ github.ref }}-${{ github.sha }}
-  cancel-in-progress: true
-
-jobs:
-  ce_job_pre_check:
-    runs-on: ubuntu-latest
-    env:
-      COMPILE_BRANCH: ${{ vars.COMPILE_BRANCH }}
-      CE_COMPILE_SELECTION: ${{ vars.CE_COMPILE_SELECTION }}
-      COMPILE_USE_PADDLE_WHL_URL_MAPPINGS: ${{ vars.COMPILE_USE_PADDLE_WHL_URL_MAPPINGS }}
-    outputs:
-      branch_match: ${{ steps.set_output.outputs.branch_match }}
-      compile_use_paddle_whl_url: ${{ steps.set_output.outputs.compile_use_paddle_whl_url }}
-      sm8689_match: ${{ steps.set_output.outputs.sm8689_match }}
-      sm8090_match: ${{ steps.set_output.outputs.sm8090_match }}
-
-    steps:
-      - name: Set Version
-        id: set_output
-        env:
-          COMPILE_BRANCH: ${{ env.COMPILE_BRANCH }}
-          CE_COMPILE_SELECTION: ${{ env.CE_COMPILE_SELECTION }}
-          COMPILE_USE_PADDLE_WHL_URL_MAPPINGS: ${{ env.COMPILE_USE_PADDLE_WHL_URL_MAPPINGS }}
-          GITHUB_REF_NAME: ${{ github.ref_name }}
-        run: |
-          # 选择要触发编译任务的分支 done
-          # 选择指定分支要编译的任务 8090或者8689
-          # 指定分支编译要使用的Paddle的安装包,默认使用nightly最新的
-
-          IFS=',' read -ra BRANCHES <<< "$COMPILE_BRANCH"
-          MATCH=false
-          for b in "${BRANCHES[@]}"; do
-            if [[ "$b" == "${GITHUB_REF_NAME}" ]]; then
-              MATCH=true
-              break
-            fi
-          done
-          echo "branch_match=$MATCH" >> $GITHUB_OUTPUT
-
-          # 通过变量CE_COMPILE_SELECTION中的映射关系,决定分支是编译sm8090还是sm8689
-          for pair in $(echo "$CE_COMPILE_SELECTION" | tr ';' ' '); do
-            branch=$(echo "$pair" | cut -d',' -f1)
-            compile_task_list=$(echo "$pair" | cut -d',' -f2)
-
-            if [[ "$branch" == "$GITHUB_REF_NAME" ]]; then
-
-              # 判断里面是否包含 sm8090 或 sm8689
-              if [[ "$compile_task_list" == *"sm8090"* ]]; then
-                echo "sm8090_match=true" >> $GITHUB_OUTPUT
-              fi
-              if [[ "$compile_task_list" == *"sm8689"* ]]; then
-                echo "sm8689_match=true" >> $GITHUB_OUTPUT
-              fi
-              break
-            fi
-          done
-
-          # 通过变量COMPILE_USE_PADDLE_WHL_URL_MAPPINGS中的映射关系,决定是否是安装指定版本的Paddle还是直接安装URL
-          for pair in $(echo $COMPILE_USE_PADDLE_WHL_URL_MAPPINGS | tr ';' ' '); do
-            branch=$(echo "$pair" | cut -d',' -f1)
-            paddle_whl_url=$(echo "$pair" | cut -d',' -f2)
-            if [[ "$branch" == "${{ github.ref_name }}" ]]; then
-              FOUND_PADDLE_URL="$paddle_whl_url"
-              echo "compile_use_paddle_whl_url=${FOUND_PADDLE_URL}" >> $GITHUB_OUTPUT
-              break
-            fi
-          done
-
-  print_ce_job_pre_check_outputs:
-    runs-on: ubuntu-latest
-    needs: ce_job_pre_check
-    steps:
-      - name: Print outputs as JSON
-        run: |
-          echo '${{ toJSON(needs.ce_job_pre_check.outputs) }}'
-
-
-  clone:
-    environment: CodeSync
-    name: FD-Clone-Linux
-    runs-on: ubuntu-latest
-    needs: ce_job_pre_check
-    if: ${{ needs.ce_job_pre_check.outputs.branch_match == 'true' }}
-    outputs:
-      repo_archive_url: ${{ steps.set_output.outputs.repo_archive_url }}
-    steps:
-      - name: Clone FastDeploy
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event_name == 'pull_request'
-                && github.event.pull_request.base.ref
-                || github.ref_name }}
-          submodules: 'recursive'
-          fetch-depth: 1000
-
-      - name: Python Setup
-        uses: actions/setup-python@v5
-        with:
-          python-version: '3.10'
-      - name: Code Info Show and Upload
-        id: set_output
-        env:
-          AK: ${{ secrets.BOS_AK }}
-          SK: ${{ secrets.BOS_SK }}
-        run: |
-          git config --unset http.https://github.com/.extraheader
-          git submodule foreach --recursive sh -c "git config --local --unset-all 'http.https://github.com/.extraheader'"
-          git submodule foreach --recursive sh -c "git config remote.origin.fetch '+refs/heads/*:refs/remotes/origin/*'"
-          echo "Current HEAD Log:"
-          git log --oneline -n 5
-          ls
-          cd ..
-          tar -zcf FastDeploy.tar.gz FastDeploy
-          commit_id=${{ github.sha }}
-          branch_name=${{ github.ref_name }}
-          target_path=paddle-qa/BRANCH/FastDeploy/${branch_name}/${commit_id}
-          wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
-          push_file=$(realpath bos_tools.py)
-          python -m pip install bce-python-sdk==0.9.29
-          ls
-          python ${push_file} FastDeploy.tar.gz ${target_path}
-          target_path_stripped="${target_path#paddle-qa/}"
-          REPO_ARCHIVE_URL=https://paddle-qa.bj.bcebos.com/${target_path_stripped}/FastDeploy.tar.gz
-          echo "repo_archive_url=${REPO_ARCHIVE_URL}" >> $GITHUB_OUTPUT
-
-  resultshow:
-    name: Show Code Archive Output
-    needs: clone
-    runs-on: ubuntu-latest
-    steps:
-      - name: Print wheel path
-        run: |
-          echo "The code archive is located at: ${{ needs.clone.outputs.repo_archive_url }}"
-
-  build_sm8090:
-    name: BUILD_SM8090
-    needs: [clone, ce_job_pre_check]
-    if: ${{ needs.ce_job_pre_check.outputs.sm8090_match == 'true' }}
-    uses: ./.github/workflows/_build_linux.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      COMPILE_ARCH: "80,90"
-      WITH_NIGHTLY_BUILD: OFF
-      FD_VERSION: 0.0.0
-      PADDLE_WHL_URL: ${{ needs.ce_job_pre_check.outputs.compile_use_paddle_whl_url }}
-
-  build_sm8689:
-    name: BUILD_SM8689
-    needs: [clone, ce_job_pre_check]
-    if: ${{ needs.ce_job_pre_check.outputs.sm8689_match == 'true' }}
-    uses: ./.github/workflows/_build_linux.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      COMPILE_ARCH: "86,89"
-      WITH_NIGHTLY_BUILD: OFF
-      FD_VERSION: 0.0.0
-      PADDLE_WHL_URL: ${{ needs.ce_job_pre_check.outputs.compile_use_paddle_whl_url }}
-
-  ce_upload_sm8090:
-    environment: CodeSync
-    name: CE_UPLOAD
-    needs: build_sm8090
-    runs-on: ubuntu-latest
-    env:
-      AK: ${{ secrets.BOS_AK }}
-      SK: ${{ secrets.BOS_SK }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
-      COMPILE_ARCH: "80,90"
-    steps:
-      - uses: actions/setup-python@v5
-        with:
-          python-version: '3.10'
-      - name: Wheel Info Show and Upload
-        run: |
-          echo "The wheel is located at: ${{ needs.build_sm8090.outputs.wheel_path }}"
-          wget -q --no-check-certificate ${{ needs.build_sm8090.outputs.wheel_path }}
-          filename=$(basename ${{ needs.build_sm8090.outputs.wheel_path }})
-
-          commit_id=${{ github.sha }}
-          branch_name=${{ github.ref_name }}
-          target_path=paddle-qa/paddle-pipeline/FastDeploy_ActionCE/SM${COMPILE_ARCH//,/_}/${branch_name}/${commit_id}
-
-          wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
-          push_file=$(realpath bos_tools.py)
-          python -m pip install bce-python-sdk==0.9.29
-          ls
-          python ${push_file} ${filename} ${target_path}
-          target_path_stripped="${target_path#paddle-qa/}"
-          WHEEL_PATH=https://paddle-qa.bj.bcebos.com/${target_path_stripped}/${fd_wheel_name}
-          echo "commit wheel url is ${WHEEL_PATH}"
-
-          target_path_latest=paddle-qa/paddle-pipeline/FastDeploy_ActionCE/SM${COMPILE_ARCH//,/_}/${branch_name}/latest
-          python ${push_file} ${filename} ${target_path_latest}
-          target_path_stripped_latest="${target_path_latest#paddle-qa/}"
-          WHEEL_PATH_LATEST=https://paddle-qa.bj.bcebos.com/${target_path_stripped_latest}/${fd_wheel_name}
-          echo "latest wheel url is ${WHEEL_PATH_LATEST}"
-
-  ce_upload_sm8689:
-    environment: CodeSync
-    name: CE_UPLOAD
-    needs: build_sm8689
-    runs-on: ubuntu-latest
-    env:
-      AK: ${{ secrets.BOS_AK }}
-      SK: ${{ secrets.BOS_SK }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8689.outputs.wheel_path }}
-      COMPILE_ARCH: "86,89"
-    steps:
-      - uses: actions/setup-python@v5
-        with:
-          python-version: '3.10'
-      - name: Wheel Info Show and Upload
-        run: |
-          echo "The wheel is located at: ${{ needs.build_sm8090.outputs.wheel_path }}"
-          wget -q --no-check-certificate ${{ needs.build_sm8090.outputs.wheel_path }}
-          filename=$(basename ${{ needs.build_sm8090.outputs.wheel_path }})
-
-          commit_id=${{ github.sha }}
-          branch_name=${{ github.ref_name }}
-          target_path=paddle-qa/paddle-pipeline/FastDeploy_ActionCE/SM${COMPILE_ARCH//,/_}/${branch_name}/${commit_id}
-
-          wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
-          push_file=$(realpath bos_tools.py)
-          python -m pip install bce-python-sdk==0.9.29
-          ls
-          python ${push_file} ${filename} ${target_path}
-          target_path_stripped="${target_path#paddle-qa/}"
-          WHEEL_PATH=https://paddle-qa.bj.bcebos.com/${target_path_stripped}/${fd_wheel_name}
-          echo "commit wheel url is ${WHEEL_PATH}"
-
-          target_path_latest=paddle-qa/paddle-pipeline/FastDeploy_ActionCE/SM${COMPILE_ARCH//,/_}/${branch_name}/latest
-          python ${push_file} ${filename} ${target_path_latest}
-          target_path_stripped_latest="${target_path_latest#paddle-qa/}"
-          WHEEL_PATH_LATEST=https://paddle-qa.bj.bcebos.com/${target_path_stripped_latest}/${fd_wheel_name}
-          echo "latest wheel url is ${WHEEL_PATH_LATEST}"

.github/workflows/check-bypass.yml

@@ -1,51 +0,0 @@
-on:
-  workflow_call:
-    inputs:
-      workflow-name:
-        required: true
-        type: string
-    secrets:
-      github-token:
-        required: true
-    outputs:
-      can-skip:
-        description: "Whether the workflow can be skipped."
-        value: ${{ jobs.check-bypass.outputs.can-skip }}
-
-jobs:
-  check-bypass:
-    name: Check bypass
-    runs-on: ubuntu-latest
-    permissions:
-      contents: read
-    env:
-      CI_TEAM_MEMBERS: '["yuanlehome","YuanRisheng","Jiang-Jia-Jun","DDDivano","XieYunshen"]'
-    outputs:
-      can-skip: ${{ steps.check-bypass.outputs.can-skip }}
-    steps:
-      - name: Cleanup
-        run: |
-          rm -rf * .[^.]*
-
-      - id: check-bypass
-        name: Check Bypass
-        uses: PFCCLab/ci-bypass@v1
-        with:
-          github-token: ${{ secrets.github-token }}
-          non-pull-request-event-strategy: 'never-skipped'
-          type: 'composite'
-          composite-rule: |
-            {
-              "any": [
-                {
-                  "type": "labeled",
-                  "label": ["skip-ci: ${{ inputs.workflow-name }}", "skip-ci: all"],
-                  "username": ${{ env.CI_TEAM_MEMBERS }}
-                },
-                {
-                  "type": "commented",
-                  "comment-pattern": [".*/skip-ci ${{ inputs.workflow-name }}.*", ".*/skip-ci all.*"],
-                  "username": ${{ env.CI_TEAM_MEMBERS }}
-                }
-              ]
-            }

.github/workflows/ci.yml

@@ -1,4 +1,4 @@
-name: CI_ILUVATAR
+name: CI
 
 on:
   pull_request:
@@ -6,13 +6,12 @@ on:
   workflow_dispatch:
 
 concurrency:
-  group: ${{ github.event.pull_request.number }}-iluvatar-ci
+  group: ${{ github.event.pull_request.number }}
   cancel-in-progress: true
 
 jobs:
-  CI_ILUVATAR:
-    runs-on:
-      group: IXUCA
+  build:
+    runs-on: [self-hosted, GPU-L20-4Card]
     steps:
       - name: Print current runner name
         run: |
@@ -23,7 +22,7 @@ jobs:
 
       - name: Code Checkout
         env:
-          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/device/paddle-ixuca:latest
+          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddle:fastdeploy-ciuse-cuda126
         run: |
           REPO="https://github.com/${{ github.repository }}.git"
           FULL_REPO="${{ github.repository }}"
@@ -52,7 +51,7 @@ jobs:
 
       - name: Run CI unittest
         env:
-          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/device/paddle-ixuca:latest
+          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddle:fastdeploy-ciuse-cuda126
         run: |
           runner_name="${{ runner.name }}"
           last_char="${runner_name: -1}"
@@ -60,7 +59,7 @@ jobs:
           if [[ "$last_char" =~ [0-3] ]]; then
             gpu_id="$last_char"
           else
-            gpu_id="0"
+            gpu_id="0"  
           fi
           FD_API_PORT=$((9180 + gpu_id * 100))
           FD_ENGINE_QUEUE_PORT=$((9150 + gpu_id * 100))
@@ -68,18 +67,17 @@ jobs:
 
           PARENT_DIR=$(dirname "$WORKSPACE")
           echo "PARENT_DIR:$PARENT_DIR"
-          docker run --rm --net=host --pid=host --cap-add=ALL --privileged --shm-size=64G  \
-          -v /usr/src:/usr/src -v /lib/modules:/lib/modules -v /dev:/dev \
-          -v $(pwd):/workspace -w /workspace \
-          -v "/data1/fastdeploy:/data1/fastdeploy" \
-          -e "MODEL_PATH=/ssd3/model" \
-          -e "http_proxy=$(git config --global --get http.proxy)" \
-          -e "https_proxy=$(git config --global --get https.proxy)" \
+          docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
+          -v "/ssd4/GithubActions/gitconfig:/etc/gitconfig:ro" \
+          -v "/ssd4/GithubActions/ModelData:/ModelData:ro" \
+          -v "/ssd4/GithubActions/CacheDir:/root/.cache" \
+          -v "/ssd4/GithubActions/ConfigDir:/root/.config" \
+          -e "MODEL_PATH=/ModelData" \
           -e "FD_API_PORT=${FD_API_PORT}" \
           -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
           -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
-           ${docker_image} /bin/bash -c "
+          --gpus device=${gpu_id} ${docker_image} /bin/bash -c "
           git config --global --add safe.directory /workspace/FastDeploy
           cd FastDeploy
-          bash scripts/run_ci_iluvatar.sh
-          "
+          bash scripts/run_ci.sh
+          "

.github/workflows/ci_gcu.yml

@@ -1,98 +0,0 @@
-name: CI_GCU
-
-on:
-  pull_request:
-    branches:
-      - develop
-      - 'release/*'
-  workflow_dispatch:
-
-concurrency:
-  group: ${{ github.event.pull_request.number }}-gcu-ci
-  cancel-in-progress: true
-
-jobs:
-  CI_GCU:
-    runs-on:
-      group: GCU
-    steps:
-      - name: Print current runner name
-        run: |
-          echo "Current runner name: ${{ runner.name }}"
-
-      - name: Code Checkout
-        env:
-          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/device/paddle-gcu:topsrider3.5.102-ubuntu20-x86_64-gcc84
-        run: |
-          REPO="https://github.com/${{ github.repository }}.git"
-          FULL_REPO="${{ github.repository }}"
-          REPO_NAME="${FULL_REPO##*/}"
-          BASE_BRANCH="${{ github.base_ref }}"
-          # Clean the repository directory before starting
-          docker run --rm --net=host -v $(pwd):/workspace \
-          -v ${{ github.workspace }}/../../..:${{ github.workspace }}/../../.. \
-          -w /workspace \
-          -e "REPO_NAME=${REPO_NAME}" \
-          -e "BASE_BRANCH=${BASE_BRANCH}" \
-          ${docker_image} /bin/bash -c '
-            if [ -d ${REPO_NAME} ]; then
-              echo "Directory ${REPO_NAME} exists, removing it..."
-              rm -rf ${REPO_NAME}
-            fi
-          '
-          git config --global user.name "FastDeployCI"
-          git config --global user.email "fastdeploy_ci@example.com"
-          source ${{ github.workspace }}/../../../proxy
-          git clone ${REPO} ${REPO_NAME} -b ${BASE_BRANCH}
-          cd FastDeploy
-          if [ "${{ github.event_name }}" = "pull_request" ]; then
-            git fetch origin pull/${{ github.event.pull_request.number }}/head:pr/${{ github.event.pull_request.number }}
-            git merge pr/${{ github.event.pull_request.number }}
-            git log -n 3 --oneline
-          else
-            git checkout ${{ github.sha }}
-            git log -n 3 --oneline
-          fi
-          echo "Copy models..."
-          sudo mkdir -p ci_models && sudo cp -r /work/deps/ERNIE-4.5-21B-A3B-Paddle ci_models
-          echo "Copy models done."
-
-      - name: Run CI unittest
-        env:
-          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/device/paddle-gcu:topsrider3.5.102-ubuntu20-x86_64-gcc84
-        run: |
-          runner_name="${{ runner.name }}"
-          last_char="${runner_name: -1}"
-
-          if [[ "$last_char" =~ [0-3] ]]; then
-            gcu_id="$last_char"
-          else
-            gcu_id="0"
-          fi
-          FD_API_PORT=$((9180 + gcu_id * 100))
-          FD_ENGINE_QUEUE_PORT=$((9150 + gcu_id * 100))
-          FD_METRICS_PORT=$((9170 + gcu_id * 100))
-
-          PARENT_DIR=$(dirname "$WORKSPACE")
-          echo "PARENT_DIR:$PARENT_DIR"
-          echo "Install drivers..."
-          cd /work/deps
-          sudo bash TopsRider_i3x_*_deb_amd64.run --driver --no-auto-load -y
-          cd -
-          echo "Create docker..."
-          docker run --rm --network=host --ipc=host --privileged \
-          -v $(pwd):/workspace \
-          -v /home:/home \
-          -v /work:/work \
-          -w /workspace \
-          -e "MODEL_PATH=./ci_models" \
-          -e "http_proxy=$(git config --global --get http.proxy)" \
-          -e "https_proxy=$(git config --global --get https.proxy)" \
-          -e "FD_API_PORT=${FD_API_PORT}" \
-          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
-          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
-          ${docker_image} /bin/bash -c "
-          git config --global --add safe.directory /workspace/FastDeploy
-          cd FastDeploy
-          bash scripts/run_ci_gcu.sh
-          "

.github/workflows/ci_xpu.yml

@@ -1,89 +0,0 @@
-name: CI_XPU
-
-on:
-  pull_request:
-    branches:
-      - develop
-      - 'release/*'
-      - 'feature/*'
-  workflow_dispatch:
-
-concurrency:
-  group: ${{ github.event.pull_request.number }}-xpu-ci
-  cancel-in-progress: true
-
-jobs:
-  CI_XPU:
-    runs-on: [self-hosted, XPU-P800-8Card]
-    steps:
-      - name: Print current runner name
-        run: |
-          echo "Current runner name: ${{ runner.name }}"
-      # Because the system version is lower than 2.23, the checkout cannot be used.
-      # - name: Checkout code
-      #   uses: actions/checkout@v4
-
-      - name: Code Checkout
-        env:
-          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/fastdeploy-xpu:2.1.0
-        run: |
-          REPO="https://github.com/${{ github.repository }}.git"
-          FULL_REPO="${{ github.repository }}"
-          REPO_NAME="${FULL_REPO##*/}"
-          BASE_BRANCH="${{ github.base_ref }}"
-          # Clean the repository directory before starting
-          docker run --rm --net=host -v $(pwd):/workspace -w /workspace \
-          -e "REPO_NAME=${REPO_NAME}" \
-          -e "BASE_BRANCH=${BASE_BRANCH}" \
-          ${docker_image} /bin/bash -c '
-            if [ -d ${REPO_NAME} ]; then
-              echo "Directory ${REPO_NAME} exists, removing it..."
-              rm -rf ${REPO_NAME}
-            fi
-          '
-          git config --global user.name "FastDeployCI"
-          git config --global user.email "fastdeploy_ci@example.com"
-          git clone ${REPO} ${REPO_NAME} -b ${BASE_BRANCH}
-          cd FastDeploy
-          if [ "${{ github.event_name }}" = "pull_request" ]; then
-            git fetch origin pull/${{ github.event.pull_request.number }}/head:pr/${{ github.event.pull_request.number }}
-            git merge pr/${{ github.event.pull_request.number }}
-            git log -n 3 --oneline
-          else
-            git checkout ${{ github.sha }}
-            git log -n 3 --oneline
-          fi
-
-      - name: Run CI unittest
-        env:
-          docker_image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/fastdeploy-xpu:2.1.0
-        run: |
-          runner_name="${{ runner.name }}"
-          last_char="${runner_name: -1}"
-
-          if [[ "$last_char" =~ [0-3] ]]; then
-            gpu_id="$last_char"
-          else
-            gpu_id="0"
-          fi
-          FD_API_PORT=$((9180 + gpu_id * 100))
-          FD_ENGINE_QUEUE_PORT=$((9150 + gpu_id * 100))
-          FD_METRICS_PORT=$((9170 + gpu_id * 100))
-
-          PARENT_DIR=$(dirname "$WORKSPACE")
-          echo "PARENT_DIR:$PARENT_DIR"
-          docker run --rm --net=host --cap-add=SYS_PTRACE --privileged --shm-size=64G  \
-          -v $(pwd):/workspace -w /workspace \
-          -v "/ssd3:/ssd3" \
-          -e "MODEL_PATH=/ssd3/model" \
-          -e "http_proxy=$(git config --global --get http.proxy)" \
-          -e "https_proxy=$(git config --global --get https.proxy)" \
-          -e "no_proxy=bcebos.com,mirrors.tuna.tsinghua.edu.cn,127.0.0.1,localhost" \
-          -e "FD_API_PORT=${FD_API_PORT}" \
-          -e "FD_ENGINE_QUEUE_PORT=${FD_ENGINE_QUEUE_PORT}" \
-          -e "FD_METRICS_PORT=${FD_METRICS_PORT}" \
-           ${docker_image} /bin/bash -c "
-          git config --global --add safe.directory /workspace/FastDeploy
-          cd FastDeploy
-          bash scripts/run_ci_xpu.sh
-          "

.github/workflows/gh-pages.yml

@@ -3,6 +3,8 @@ name: Deploy GitHub Pages
 on:
   push:
     branches: [ develop ]
+  pull_request:
+    branches: [ develop ]
 
 permissions:
   contents: write
@@ -15,10 +17,8 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: 3.x
-      - run: pip install mkdocs-material mkdocs-get-deps mkdocs-material-extensions mkdocs-multilang mkdocs-static-i18n
+      - run: pip install mkdocs-material mkdocs-get-deps mkdocs-material-extensions mkdocs-multilang
       - name: Deploy to GitHub Pages
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          git remote set-url origin https://x-access-token:${{ secrets.GITHUB_TOKEN }}@github.com/${{ github.repository }}.git
-          mkdocs gh-deploy --force --remote-name origin
+        run: mkdocs gh-deploy --force --remote-name origin

.github/workflows/pr_build_and_test.yml

@@ -1,97 +0,0 @@
-name: PR Build and Test
-on:
-  pull_request:
-    types: [opened, synchronize]
-    branches: [develop, release/**, feature/**]
-permissions: read-all
-
-concurrency:
-  group: ${{ github.event.pull_request.number }}-${{ github.workflow }}
-  cancel-in-progress: true
-
-jobs:
-  clone:
-    name: FD-Clone-Linux
-    uses: ./.github/workflows/_clone_linux.yml
-
-  build:
-    name: FD-Build-Linux
-    needs: clone
-    uses: ./.github/workflows/_build_linux.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      COMPILE_ARCH: "89,90"
-      WITH_NIGHTLY_BUILD: "OFF"
-      FD_VERSION: "0.0.0"
-
-  resultshow:
-    name: Use Build Output
-    needs: build
-    runs-on: ubuntu-latest
-    steps:
-      - name: Print wheel path
-        run: |
-          echo "The built wheel is located at: ${{ needs.build.outputs.wheel_path }}"
-
-  unittest_coverage:
-    name: Run FastDeploy Unit Tests and Coverage
-    needs: [clone,build]
-    uses: ./.github/workflows/_unit_test_coverage.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
-      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-
-  logprob_test:
-    name: Run FastDeploy LogProb Tests
-    needs: [build]
-    uses: ./.github/workflows/_logprob_test_linux.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      PADDLETEST_ARCHIVE_URL: "https://xly-devops.bj.bcebos.com/PaddleTest/PaddleTest.tar.gz"
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
-      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
-
-  pre_ce_test:
-    name: Extracted partial CE model tasks to run in CI.
-    needs: [clone,build]
-    uses: ./.github/workflows/_pre_ce_test.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
-      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
-
-  base_test:
-    name: Run Base Tests
-    needs: [clone,build]
-    uses: ./.github/workflows/_base_test.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
-      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
-
-  accuracy_test:
-    name: Run Accuracy Tests
-    needs: [clone,build]
-    uses: ./.github/workflows/_accuracy_test.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
-      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
-
-  stable_test:
-    name: Run Stable Tests
-    needs: [clone,build]
-    uses: ./.github/workflows/_stable_test.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build.outputs.wheel_path }}
-      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"

.github/workflows/publish_job.yml

@@ -1,321 +0,0 @@
-name: Publish Job
-
-on:
-  workflow_dispatch:
-  schedule:
-    - cron: '0 18 * * *'   # 2:00 AM China Standard Time (UTC+8)
-  push:
-    # branches:
-    #   - develop
-    tags:
-      - '*'
-
-permissions: read-all
-
-concurrency:
-  group: ${{ github.ref }}-${{ github.sha }}
-  cancel-in-progress: true
-
-
-jobs:
-  publish_pre_check:
-    runs-on: ubuntu-latest
-    if: |
-      github.event.repository.fork == false &&
-      (
-        (github.event_name == 'schedule' && github.ref_name == 'develop') ||
-        (github.event_name == 'push' && github.ref_type == 'tag') ||
-        ((github.event_name == 'workflow_dispatch') &&
-          (github.ref_name == 'develop' || github.ref_type == 'tag'))
-      )
-    env:
-      TAG_VERSION_MAPPINGS: ${{ vars.TAG_VERSION_MAPPINGS }}
-      FD_VERSION_DEV: ${{ vars.FD_VERSION_DEV }}
-      COMPILE_USE_PADDLE_WHL_URL_MAPPINGS: ${{ vars.COMPILE_USE_PADDLE_WHL_URL_MAPPINGS }}
-    outputs:
-      compile_use_paddle_version: ${{ steps.set_output.outputs.compile_use_paddle_version }}
-      compile_continue: ${{ steps.set_output.outputs.compile_continue }}
-      fd_version: ${{ steps.set_output.outputs.fd_version }}
-      with_nightly_build: ${{ steps.set_output.outputs.with_nightly_build }}
-      compile_use_paddle_whl_url: ${{ steps.set_output.outputs.compile_use_paddle_whl_url }}
-
-    steps:
-      - name: Get tag version
-        if: github.ref_type == 'tag'
-        run: |
-          TAG_NAME="${GITHUB_REF##*/}"       # 提取 tag 名称，比如 v2.1.0
-          TAG_VERSION="${TAG_NAME#v}"            # 去掉前缀 v
-          echo "FD_VERSION=$TAG_VERSION" >> $GITHUB_ENV
-
-      - name: Check FD version to Paddle version mapping
-        if: github.ref_type == 'tag'
-        env:
-          TARGET_FD: ${{ env.FD_VERSION }}
-        run: |
-          FOUND_PADDLE=""
-          # 遍历映射
-          for pair in $(echo $TAG_VERSION_MAPPINGS | tr ';' ' '); do
-            fd=$(echo "$pair" | cut -d',' -f1)
-            paddle=$(echo "$pair" | cut -d',' -f2)
-            if [[ "$fd" == "$TARGET_FD" ]]; then
-              FOUND_PADDLE="$paddle"
-              break
-            fi
-          done
-
-          if [[ -z "$FOUND_PADDLE" ]]; then
-            echo "No Paddle version found for FD $TARGET_FD"
-          else
-            echo "FD $TARGET_FD maps to Paddle $FOUND_PADDLE"
-            echo "PADDLE_VERSION=$FOUND_PADDLE" >> $GITHUB_ENV
-          fi
-      - name: Set Version
-        id: set_output
-        env:
-          PADDLE_VERSION: ${{ env.PADDLE_VERSION }}
-          FD_VERSION: ${{ env.FD_VERSION }}
-        run: |
-          if [[ "${{ github.ref_type }}" == "tag" ]]; then
-            if [[ -z "$PADDLE_VERSION" ]]; then
-              compile_continue=false
-            else
-              compile_use_paddle_version=$PADDLE_VERSION
-              compile_continue=true
-            fi
-            fd_version=$FD_VERSION
-          fi
-          if [[ "${{ github.ref_name }}" == "develop" ]];then
-            compile_continue=true
-            compile_use_paddle_version=""
-            fd_version=${FD_VERSION_DEV}
-            with_nightly_build=ON
-          fi
-          # Todo
-          # 通过变量COMPILE_USE_PADDLE_WHL_URL_MAPPINGS中的映射关系,决定是否是安装指定版本的Paddle还是直接安装URL
-          for pair in $(echo $COMPILE_USE_PADDLE_WHL_URL_MAPPINGS | tr ';' ' '); do
-            branch=$(echo "$pair" | cut -d',' -f1)
-            paddle_whl_url=$(echo "$pair" | cut -d',' -f2)
-            if [[ "$branch" == "${{ github.ref_name }}" ]]; then
-              FOUND_PADDLE_URL="$paddle_whl_url"
-              echo "compile_use_paddle_whl_url=${FOUND_PADDLE_URL}" >> $GITHUB_OUTPUT
-              compile_continue=true
-              break
-            fi
-          done
-          echo "compile_continue=${compile_continue}" >> $GITHUB_OUTPUT
-          echo "compile_use_paddle_version=${compile_use_paddle_version}" >> $GITHUB_OUTPUT
-          echo "fd_version=${fd_version}" >> $GITHUB_OUTPUT
-          echo "with_nightly_build=${with_nightly_build:-OFF}" >> $GITHUB_OUTPUT
-
-  print_publish_pre_check_outputs:
-    runs-on: ubuntu-latest
-    needs: publish_pre_check
-    steps:
-      - name: Print outputs as JSON
-        run: |
-          echo '${{ toJSON(needs.publish_pre_check.outputs) }}'
-
-  clone:
-    environment: CodeSync
-    name: FD-Clone-Linux
-    runs-on: ubuntu-latest
-    needs: publish_pre_check
-    if: ${{ needs.publish_pre_check.outputs.compile_continue == 'true' }}
-    outputs:
-      repo_archive_url: ${{ steps.set_output.outputs.repo_archive_url }}
-    steps:
-      - name: Clone FastDeploy
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.ref_name }}
-          submodules: 'recursive'
-          fetch-depth: 1000
-
-      - name: Python Setup
-        uses: actions/setup-python@v5
-        with:
-          python-version: '3.10'
-      - name: Code Info Show and Upload
-        id: set_output
-        env:
-          AK: ${{ secrets.BOS_AK }}
-          SK: ${{ secrets.BOS_SK }}
-        run: |
-          git config --unset http.https://github.com/.extraheader
-          git submodule foreach --recursive sh -c "git config --local --unset-all 'http.https://github.com/.extraheader'"
-          git submodule foreach --recursive sh -c "git config remote.origin.fetch '+refs/heads/*:refs/remotes/origin/*'"
-          echo "Current HEAD Log:"
-          git log --oneline -n 5
-          ls
-          cd ..
-          tar -zcf FastDeploy.tar.gz FastDeploy
-          if [[ "${{ github.ref_type }}" == "tag" ]]; then
-            commit_id=${{ github.sha }}
-            tag_name=${{ github.ref_name }}
-            target_path=paddle-qa/TAG/FastDeploy/${tag_name}/${commit_id}
-          else
-            commit_id=${{ github.sha }}
-            branch_name=${{ github.ref_name }}
-            target_path=paddle-qa/BRANCH/FastDeploy/${branch_name}/${commit_id}
-          fi
-          wget  -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
-          push_file=$(realpath bos_tools.py)
-          python -m pip install bce-python-sdk==0.9.29
-          ls
-          python ${push_file} FastDeploy.tar.gz ${target_path}
-          target_path_stripped="${target_path#paddle-qa/}"
-          REPO_ARCHIVE_URL=https://paddle-qa.bj.bcebos.com/${target_path_stripped}/FastDeploy.tar.gz
-          echo "repo_archive_url=${REPO_ARCHIVE_URL}" >> $GITHUB_OUTPUT
-
-  resultshow:
-    name: Show Code Archive Output
-    needs: clone
-    runs-on: ubuntu-latest
-    steps:
-      - name: Print wheel path
-        run: |
-          echo "The code archive is located at: ${{ needs.clone.outputs.repo_archive_url }}"
-
-  build_sm8090:
-    name: BUILD_SM8090
-    needs: [clone, publish_pre_check]
-    uses: ./.github/workflows/_build_linux.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      COMPILE_ARCH: "80,90"
-      WITH_NIGHTLY_BUILD: ${{ needs.publish_pre_check.outputs.with_nightly_build }}
-      FD_VERSION: ${{ needs.publish_pre_check.outputs.fd_version }}
-      PADDLEVERSION: ${{ needs.publish_pre_check.outputs.compile_use_paddle_version }}
-      PADDLE_WHL_URL: ${{ needs.publish_pre_check.outputs.compile_use_paddle_whl_url }}
-
-  build_sm8689:
-    name: BUILD_SM8689
-    needs: [clone, publish_pre_check]
-    uses: ./.github/workflows/_build_linux.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      COMPILE_ARCH: "86,89"
-      WITH_NIGHTLY_BUILD: ${{ needs.publish_pre_check.outputs.with_nightly_build }}
-      FD_VERSION: ${{ needs.publish_pre_check.outputs.fd_version }}
-      PADDLEVERSION: ${{ needs.publish_pre_check.outputs.compile_use_paddle_version }}
-      PADDLE_WHL_URL: ${{ needs.publish_pre_check.outputs.compile_use_paddle_whl_url }}
-
-  paddle_pypi_upload_sm8090:
-    environment: PaddleSourceUpload
-    name: PADDLE_PYPI_UPLOAD_8090
-    needs: build_sm8090
-    runs-on: ubuntu-latest
-    env:
-      AK: ${{ secrets.BOS_AK }}
-      SK: ${{ secrets.BOS_SK }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
-      COMPILE_ARCH: "80,90"
-    steps:
-      - uses: actions/setup-python@v5
-        with:
-          python-version: '3.10'
-      - name: Wheel Info Show and Upload
-        if: github.ref_name == 'develop' || github.ref_type == 'tag'
-        run: |
-          echo "The wheel is located at: ${FASTDEPLOY_WHEEL_URL}"
-          wget -q --no-check-certificate ${FASTDEPLOY_WHEEL_URL}
-          filename=$(basename ${FASTDEPLOY_WHEEL_URL})
-          if [[ "${{ github.ref_name }}" == "develop" ]];then
-            target_path=paddle-whl/nightly/fastdeploy-gpu-${COMPILE_ARCH//,/_}/fastdeploy-gpu
-          elif [[ "${{ github.ref_type }}" == "tag" ]]; then
-            target_path=paddle-whl/stable/fastdeploy-gpu-${COMPILE_ARCH//,/_}/fastdeploy-gpu
-          else
-            echo "Not develop or tag, do nothing"
-          fi
-          wget -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
-          push_file=$(realpath bos_tools.py)
-          python -m pip install bce-python-sdk==0.9.29
-          ls
-          python ${push_file} ${filename} ${target_path}
-
-  paddle_pypi_upload_sm8689:
-    environment: PaddleSourceUpload
-    name: PADDLE_PYPI_UPLOAD_8689
-    needs: build_sm8689
-    runs-on: ubuntu-latest
-    env:
-      AK: ${{ secrets.BOS_AK }}
-      SK: ${{ secrets.BOS_SK }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8689.outputs.wheel_path }}
-      COMPILE_ARCH: "86,89"
-    steps:
-      - uses: actions/setup-python@v5
-        with:
-          python-version: '3.10'
-      - name: Wheel Info Show and Upload
-        if: github.ref_name == 'develop' || github.ref_type == 'tag'
-        run: |
-          echo "The wheel is located at: ${FASTDEPLOY_WHEEL_URL}"
-          wget -q --no-check-certificate ${FASTDEPLOY_WHEEL_URL}
-          filename=$(basename ${FASTDEPLOY_WHEEL_URL})
-          if [[ "${{ github.ref_name }}" == "develop" ]];then
-            target_path=paddle-whl/nightly/fastdeploy-gpu-${COMPILE_ARCH//,/_}/fastdeploy-gpu
-          elif [[ "${{ github.ref_type }}" == "tag" ]]; then
-            target_path=paddle-whl/stable/fastdeploy-gpu-${COMPILE_ARCH//,/_}/fastdeploy-gpu
-          else
-            echo "Not develop or tag, do nothing"
-          fi
-          wget -q --no-proxy --no-check-certificate https://paddle-qa.bj.bcebos.com/CodeSync/develop/PaddlePaddle/PaddleTest/tools/bos_tools.py
-          push_file=$(realpath bos_tools.py)
-          python -m pip install bce-python-sdk==0.9.29
-          ls
-          python ${push_file} ${filename} ${target_path}
-
-  unittest_coverage:
-    name: Run FastDeploy Unit Tests and Coverage
-    needs: [clone,build_sm8090]
-    uses: ./.github/workflows/_unit_test_coverage.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
-      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-
-  logprob_test:
-    name: Run FastDeploy LogProb Tests
-    needs: [build_sm8090]
-    uses: ./.github/workflows/_logprob_test_linux.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      PADDLETEST_ARCHIVE_URL: "https://xly-devops.bj.bcebos.com/PaddleTest/PaddleTest.tar.gz"
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
-      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
-
-  pre_ce_test:
-    name: Extracted partial CE model tasks to run in CI.
-    needs: [clone,build_sm8090]
-    uses: ./.github/workflows/_pre_ce_test.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
-      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
-
-  base_test:
-    name: Run Base Tests
-    needs: [clone,build_sm8090]
-    uses: ./.github/workflows/_base_test.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
-      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"
-
-  accuracy_test:
-    name: Run Accuracy Tests
-    needs: [clone,build_sm8090]
-    uses: ./.github/workflows/_accuracy_test.yml
-    with:
-      DOCKER_IMAGE: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleqa:fastdeploy-ciuse-cuda126-dailyupdate
-      FASTDEPLOY_ARCHIVE_URL: ${{ needs.clone.outputs.repo_archive_url }}
-      FASTDEPLOY_WHEEL_URL: ${{ needs.build_sm8090.outputs.wheel_path }}
-      MODEL_CACHE_DIR: "/ssd2/actions-runner/ModelData"

.gitignore

@@ -121,7 +121,7 @@ dmypy.json
 FETCH_HEAD
 
 #log
-log/
+log*/
 
 checkpoints/
 checkpoints_origin/
@@ -156,23 +156,9 @@ nohup.out
 custom_ops/gpu_ops/fp8_deep_gemm/deep_gemm/include/cutlass
 custom_ops/gpu_ops/fp8_deep_gemm/deep_gemm/include/cute
 
-#marlin_kernel
-custom_ops/gpu_ops/moe/moe_wna16_marlin_utils/kernel_*.cu
-
-#machete_kernel
-custom_ops/gpu_ops/machete/generated
-
 # buff
 custom_ops/tmp*
 
 build
 
 .ccls-cache
-
-third_party
-
-custom_ops/gpu_ops/w4afp8_gemm/w4afp8_gemm_*.cu
-custom_ops/gpu_ops/w4afp8_gemm/w4afp8_gemm_template.h
-
-custom_ops/gpu_ops/wfp8afp8_sparse_gemm/wfp8Afp8_sparse_gemm_*.cu
-custom_ops/gpu_ops/wfp8afp8_sparse_gemm/wfp8Afp8_sparse_gemm_template.h

.pre-commit-config.yaml

@@ -3,30 +3,20 @@ default_install_hook_types:
   - commit-msg
 default_stages:
   - pre-commit # Run locally
-  - commit-msg
 #   - manual # Run in CI
 repos:
--   repo: https://github.com/psf/black.git
-    rev: 25.1.0
-    hooks:
-    -   id: black
-        files: \.(py|pyi)$
-        additional_dependencies: [toml]
-# 自动排序
--   repo: https://github.com/PyCQA/isort
-    rev: 5.11.5
-    hooks:
-    -   id: isort
--   repo: https://github.com/PyCQA/flake8
-    rev: 7.0.0
-    hooks:
-    -   id: flake8
+# 格式化
+- repo: https://github.com/google/yapf
+  rev: v0.43.0
+  hooks:
+  - id: yapf
+    args: [--in-place, --verbose]
 # 代码检查
 - repo: https://github.com/astral-sh/ruff-pre-commit
   rev: v0.11.7
   hooks:
   - id: ruff
-    args: [--output-format, github, --fix, --line-length=120, --config, pyproject.toml]
+    args: [--output-format, github, --fix, --line-length=120]
 # # 拼写检查
 # - repo: https://github.com/codespell-project/codespell
 #   rev: v2.4.1
@@ -34,13 +24,26 @@ repos:
 #   - id: codespell
 #     additional_dependencies: ['tomli']
 #     args: ['--toml', 'pyproject.toml']
+# 自动排序
+- repo: https://github.com/PyCQA/isort
+  rev: 6.0.1
+  hooks:
+  - id: isort
+# # 格式化
+# - repo: https://github.com/pre-commit/mirrors-clang-format
+#   rev: v20.1.3
+#   hooks:
+#   - id: clang-format
+#     # exclude: '.*'
+#     types_or: [c++, cuda]
+#     args: [--style=file, --verbose]
 
 # markdown
 - repo: https://github.com/jackdewinter/pymarkdown
   rev: v0.9.29
   hooks:
   - id: pymarkdown
-    args: ["-d", "MD029,MD031", fix]
+    args: [fix]
 - repo: https://github.com/pre-commit/pre-commit-hooks
   rev: v5.0.0
   hooks:

README.md

@@ -1,4 +1,3 @@
-English | [简体中文](README_CN.md)
 <p align="center">
   <a href="https://github.com/PaddlePaddle/FastDeploy/releases"><img src="https://github.com/user-attachments/assets/42b0039f-39e3-4279-afda-6d1865dfbffb" width="500"></a>
 </p>
@@ -9,26 +8,20 @@ English | [简体中文](README_CN.md)
     <a href="https://github.com/PaddlePaddle/FastDeploy/commits"><img src="https://img.shields.io/github/commit-activity/m/PaddlePaddle/FastDeploy?color=3af"></a>
     <a href="https://github.com/PaddlePaddle/FastDeploy/issues"><img src="https://img.shields.io/github/issues/PaddlePaddle/FastDeploy?color=9cc"></a>
     <a href="https://github.com/PaddlePaddle/FastDeploy/stargazers"><img src="https://img.shields.io/github/stars/PaddlePaddle/FastDeploy?color=ccf"></a>
-
 </p>
 
 <p align="center">
-     <a href="https://trendshift.io/repositories/4046" target="_blank"><img src="https://trendshift.io/api/badge/repositories/4046" alt="PaddlePaddle%2FFastDeploy | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a></br>
     <a href="https://paddlepaddle.github.io/FastDeploy/get_started/installation/nvidia_gpu/"><b> Installation </b></a>
     |
     <a href="https://paddlepaddle.github.io/FastDeploy/get_started/quick_start"><b> Quick Start </b></a>
     |
     <a href="https://paddlepaddle.github.io/FastDeploy/supported_models/"><b> Supported Models </b></a>
-
 </p>
 
 --------------------------------------------------------------------------------
-# FastDeploy : Inference and Deployment Toolkit for LLMs and VLMs based on PaddlePaddle
+# FastDeploy 2.0: Inference and Deployment Toolkit for LLMs and VLMs based on PaddlePaddle
 
 ## News
-**[2025-08] 🔥 Released FastDeploy v2.1:** A brand-new KV Cache scheduling strategy has been introduced, and expanded support for PD separation and CUDA Graph across more models. Enhanced hardware support has been added for platforms like Kunlun and Hygon, along with comprehensive optimizations to improve the performance of both the service and inference engine.
-
-**[2025-07] The FastDeploy 2.0 Inference Deployment Challenge is now live!** Complete the inference deployment task for the ERNIE 4.5 series open-source models to win official FastDeploy 2.0 merch and generous prizes! 🎁 You're welcome to try it out and share your feedback! 📌[Sign up here](https://www.wjx.top/vm/meSsp3L.aspx#) 📌[Event details](https://github.com/PaddlePaddle/FastDeploy/discussions/2728)
 
 **[2025-06] 🔥 Released FastDeploy v2.0:** Supports inference and deployment for ERNIE 4.5. Furthermore, we open-source an industrial-grade PD disaggregation with context caching, dynamic role switching for effective resource utilization to further enhance inference performance for MoE models.
 
@@ -50,15 +43,14 @@ English | [简体中文](README_CN.md)
 
 ## Installation
 
-FastDeploy supports inference deployment on **NVIDIA GPUs**, **Kunlunxin XPUs**, **Iluvatar GPUs**, **Enflame GCUs**, **Hygon DCUs** and other hardware. For detailed installation instructions:
+FastDeploy supports inference deployment on **NVIDIA GPUs**, **Kunlunxin XPUs**, **Iluvatar GPUs**, **Enflame GCUs**, and other hardware. For detailed installation instructions:
 
 - [NVIDIA GPU](./docs/get_started/installation/nvidia_gpu.md)
 - [Kunlunxin XPU](./docs/get_started/installation/kunlunxin_xpu.md)
 - [Iluvatar GPU](./docs/get_started/installation/iluvatar_gpu.md)
 - [Enflame GCU](./docs/get_started/installation/Enflame_gcu.md)
-- [Hygon DCU](./docs/get_started/installation/hygon_dcu.md)
 
-**Note:** We are actively working on expanding hardware support. Additional hardware platforms including Ascend NPU and MetaX GPU are currently under development and testing. Stay tuned for updates!
+**Note:** We are actively working on expanding hardware support. Additional hardware platforms including Ascend NPU, Hygon DCU, and MetaX GPU are currently under development and testing. Stay tuned for updates!
 
 ## Get Started
 
@@ -69,19 +61,18 @@ Learn how to use FastDeploy through our documentation:
 - [Offline Inference Development](./docs/offline_inference.md)
 - [Online Service Deployment](./docs/online_serving/README.md)
 - [Full Supported Models List](./docs/supported_models.md)
-- [Best Practices](./docs/best_practices/README.md)
 
 ## Supported Models
 
 | Model | Data Type | PD Disaggregation | Chunked Prefill | Prefix Caching |  MTP | CUDA Graph | Maximum Context Length |
 |:--- | :------- | :---------- | :-------- | :-------- | :----- | :----- | :----- |
-|ERNIE-4.5-300B-A47B | BF16/WINT4/WINT8/W4A8C8/WINT2/FP8 | ✅| ✅ | ✅|✅| ✅ |128K |
-|ERNIE-4.5-300B-A47B-Base| BF16/WINT4/WINT8 | ✅| ✅ | ✅|❌| ✅ | 128K |
+|ERNIE-4.5-300B-A47B | BF16/WINT4/WINT8/W4A8C8/WINT2/FP8 | ✅| ✅ | ✅|✅(WINT4)| WIP |128K |
+|ERNIE-4.5-300B-A47B-Base| BF16/WINT4/WINT8 | ✅| ✅ | ✅|✅(WINT4)| WIP | 128K |
 |ERNIE-4.5-VL-424B-A47B | BF16/WINT4/WINT8 | WIP | ✅ | WIP | ❌ | WIP |128K |
 |ERNIE-4.5-VL-28B-A3B | BF16/WINT4/WINT8 | ❌ | ✅ | WIP | ❌ | WIP |128K |
-|ERNIE-4.5-21B-A3B | BF16/WINT4/WINT8/FP8  |  ❌ |  ✅ |  ✅ | ✅ | ✅|128K |
-|ERNIE-4.5-21B-A3B-Base | BF16/WINT4/WINT8/FP8  |  ✅ |  ✅ |  ✅ | ❌  | ✅|128K |
-|ERNIE-4.5-0.3B | BF16/WINT8/FP8  |  ✅ |  ✅ |  ✅ | ❌ | ✅| 128K |
+|ERNIE-4.5-21B-A3B | BF16/WINT4/WINT8/FP8  |  ❌ |  ✅ |  ✅ | WIP | ✅|128K |
+|ERNIE-4.5-21B-A3B-Base | BF16/WINT4/WINT8/FP8  |  ❌ |  ✅ |  ✅ | WIP | ✅|128K |
+|ERNIE-4.5-0.3B | BF16/WINT8/FP8  |  ❌ |  ✅ |  ✅ | ❌ | ✅| 128K |
 
 ## Advanced Usage
 

README_CN.md

@@ -1,94 +0,0 @@
-[English](README.md) | 简体中文
-<p align="center">
-  <a href="https://github.com/PaddlePaddle/FastDeploy/releases"><img src="https://github.com/user-attachments/assets/42b0039f-39e3-4279-afda-6d1865dfbffb" width="500"></a>
-</p>
-<p align="center">
-    <a href=""><img src="https://img.shields.io/badge/python-3.10-aff.svg"></a>
-    <a href=""><img src="https://img.shields.io/badge/os-linux-pink.svg"></a>
-    <a href="https://github.com/PaddlePaddle/FastDeploy/graphs/contributors"><img src="https://img.shields.io/github/contributors/PaddlePaddle/FastDeploy?color=9ea"></a>
-    <a href="https://github.com/PaddlePaddle/FastDeploy/commits"><img src="https://img.shields.io/github/commit-activity/m/PaddlePaddle/FastDeploy?color=3af"></a>
-    <a href="https://github.com/PaddlePaddle/FastDeploy/issues"><img src="https://img.shields.io/github/issues/PaddlePaddle/FastDeploy?color=9cc"></a>
-    <a href="https://github.com/PaddlePaddle/FastDeploy/stargazers"><img src="https://img.shields.io/github/stars/PaddlePaddle/FastDeploy?color=ccf"></a>
-
-</p>
-
-<p align="center">
-     <a href="https://trendshift.io/repositories/4046" target="_blank"><img src="https://trendshift.io/api/badge/repositories/4046" alt="PaddlePaddle%2FFastDeploy | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a></br>
-    <a href="https://paddlepaddle.github.io/FastDeploy/zh/get_started/installation/nvidia_gpu/"><b> 安装指导 </b></a>
-    |
-    <a href="https://paddlepaddle.github.io/FastDeploy/zh/get_started/quick_start"><b> 快速入门 </b></a>
-    |
-    <a href="https://paddlepaddle.github.io/FastDeploy/zh/supported_models/"><b> 支持模型列表 </b></a>
-
-</p>
-
---------------------------------------------------------------------------------
-# FastDeploy ：基于飞桨的大语言模型与视觉语言模型推理部署工具包
-
-## 最新活动
-**[2025-08] 🔥 FastDeploy v2.1 全新发布:** 全新的KV Cache调度策略，更多模型支持PD分离和CUDA Graph，昆仑、海光等更多硬件支持增强，全方面优化服务和推理引擎的性能。
-
-**[2025-07] 《FastDeploy2.0推理部署实测》专题活动已上线!** 完成文心4.5系列开源模型的推理部署等任务，即可获得骨瓷马克杯等FastDeploy2.0官方周边及丰富奖金！🎁 欢迎大家体验反馈～ 📌[报名地址](https://www.wjx.top/vm/meSsp3L.aspx#)   📌[活动详情](https://github.com/PaddlePaddle/FastDeploy/discussions/2728)
-
-## 关于
-
-**FastDeploy** 是基于飞桨（PaddlePaddle）的大语言模型（LLM）与视觉语言模型（VLM）推理部署工具包，提供**开箱即用的生产级部署方案**，核心技术特性包括：
-
-- 🚀 **负载均衡式PD分解**：工业级解决方案，支持上下文缓存与动态实例角色切换，在保障SLO达标和吞吐量的同时优化资源利用率
-- 🔄 **统一KV缓存传输**：轻量级高性能传输库，支持智能NVLink/RDMA选择
-- 🤝 **OpenAI API服务与vLLM兼容**：单命令部署，兼容[vLLM](https://github.com/vllm-project/vllm/)接口
-- 🧮 **全量化格式支持**：W8A16、W8A8、W4A16、W4A8、W2A16、FP8等
-- ⏩ **高级加速技术**：推测解码、多令牌预测（MTP）及分块预填充
-- 🖥️ **多硬件支持**：NVIDIA GPU、昆仑芯XPU、海光DCU、昇腾NPU、天数智芯GPU、燧原GCU、沐曦GPU等
-
-## 要求
-
-- 操作系统: Linux
-- Python: 3.10 ~ 3.12
-
-## 安装
-
-FastDeploy 支持在**英伟达（NVIDIA）GPU**、**昆仑芯（Kunlunxin）XPU**、**天数（Iluvatar）GPU**、**燧原（Enflame）GCU**、**海光（Hygon）DCU** 以及其他硬件上进行推理部署。详细安装说明如下：
-
-- [英伟达 GPU](./docs/zh/get_started/installation/nvidia_gpu.md)
-- [昆仑芯 XPU](./docs/zh/get_started/installation/kunlunxin_xpu.md)
-- [天数 CoreX](./docs/zh/get_started/installation/iluvatar_gpu.md)
-- [燧原 S60](./docs/zh/get_started/installation/Enflame_gcu.md)
-- [海光 DCU](./docs/zh/get_started/installation/hygon_dcu.md)
-
-**注意:** 我们正在积极拓展硬件支持范围。目前，包括昇腾（Ascend）NPU 和 沐曦（MetaX）GPU 在内的其他硬件平台正在开发测试中。敬请关注更新！
-
-## 入门指南
-
-通过我们的文档了解如何使用 FastDeploy：
-- [10分钟快速部署](./docs/zh/get_started/quick_start.md)
-- [ERNIE-4.5 部署](./docs/zh/get_started/ernie-4.5.md)
-- [ERNIE-4.5-VL 部署](./docs/zh/get_started/ernie-4.5-vl.md)
-- [离线推理](./docs/zh/offline_inference.md)
-- [在线服务](./docs/zh/online_serving/README.md)
-- [模型支持列表](./docs/zh/supported_models.md)
-- [最佳实践](./docs/zh/best_practices/README.md)
-
-## 支持模型列表
-
-| Model | Data Type | PD Disaggregation | Chunked Prefill | Prefix Caching |  MTP | CUDA Graph | Maximum Context Length |
-|:--- | :------- | :---------- | :-------- | :-------- | :----- | :----- | :----- |
-|ERNIE-4.5-300B-A47B | BF16/WINT4/WINT8/W4A8C8/WINT2/FP8 | ✅| ✅ | ✅|✅| ✅ |128K |
-|ERNIE-4.5-300B-A47B-Base| BF16/WINT4/WINT8 | ✅| ✅ | ✅|❌| ✅ | 128K |
-|ERNIE-4.5-VL-424B-A47B | BF16/WINT4/WINT8 | WIP | ✅ | WIP | ❌ | WIP |128K |
-|ERNIE-4.5-VL-28B-A3B | BF16/WINT4/WINT8 | ❌ | ✅ | WIP | ❌ | WIP |128K |
-|ERNIE-4.5-21B-A3B | BF16/WINT4/WINT8/FP8  |  ❌ |  ✅ |  ✅ | ✅ | ✅|128K |
-|ERNIE-4.5-21B-A3B-Base | BF16/WINT4/WINT8/FP8  |  ✅ |  ✅ |  ✅ | ❌  | ✅|128K |
-|ERNIE-4.5-0.3B | BF16/WINT8/FP8  |  ✅ |  ✅ |  ✅ | ❌ | ✅| 128K |
-
-## 进阶用法
-
-- [量化](./docs/zh/quantization/README.md)
-- [分离式部署](./docs/zh/features/disaggregated.md)
-- [投机解码](./docs/zh/features/speculative_decoding.md)
-- [前缀缓存](./docs/zh/features/prefix_caching.md)
-- [分块预填充](./docs/zh/features/chunked_prefill.md)
-
-## 致谢
-
-FastDeploy 依据 [Apache-2.0 开源许可证](./LICENSE). 进行授权。在开发过程中，我们参考并借鉴了 [vLLM](https://github.com/vllm-project/vllm) 的部分代码，以保持接口兼容性，在此表示衷心感谢。

benchmarks/README.md

@@ -41,10 +41,7 @@ python -m pip install -r requirements.txt
 --metric-percentiles 80,95,99,99.9,99.95,99.99：性能结果中展示的性能指标分位值
 --num-prompts 1：总计发送多少条请求
 --max-concurrency 1：压测并发数
---save-result：开启结果保存，结果文件会存入json，默认False不保存
---debug：开启debug模式，逐条打印payload和output内容，默认False
---shuffle：是否打乱数据集，默认False不打乱
---seed：打乱数据集时的随机种子，默认0
+--save-result：开启结果保存，结果文件会存入json
 ```
 
 ##### /v1/chat/completions接口压测单条数据调试
@@ -108,30 +105,3 @@ python benchmark_serving.py \
   --save-result > infer_log.txt 2>&1 &
 ```
 
-### 投机解码性能测试工具
-
-#### 使用方式：
-
-```bash
-python benchmarks/benchmark_mtp.py \
-  --host 127.0.0.1 --port 8000 \
-  --max-concurrency 16 32 64 96 --num-prompts 256 \
-  --acceptance-rate 0.8 --draft-token-steps 1 2 3 \
-  --s_itl-base-model 15.88 22.84 16.47 16.93 \
-  --dataset-name EBChat \
-  --dataset-path ./filtered_sharedgpt_2000_input_1136_output_200_fd.json
-```
-
-#### 参数说明
-
-```bash
---host：服务ip地址，用于组url
---port：服务HTTP端口，用于组url
---max-concurrency：测试并发数
---num-prompts：总计发送多少条请求
---acceptance-rate：投机解码的模拟接受率
---draft-token-steps：投机解码的步数
---s_itl-base-model：主模型的解码延迟，可由上述的性能压测工具获得，与batch-size一一对应
---dataset-name：指定数据集类，指定为"EBChat"可读取转存的FD格式数据集
---dataset-path：测试数据集路径
-```

benchmarks/backend_request_func.py

@@ -29,14 +29,13 @@ from typing import Optional
 import aiohttp
 from tqdm.asyncio import tqdm
 
+
 AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
 
 
 @dataclass
 class RequestFuncInput:
     """Input for requesting LLMs via API"""
-
-    no: int
     prompt: str
     history_QA: Optional[dict]
     hyper_parameters: dict
@@ -50,14 +49,11 @@ class RequestFuncInput:
     multi_modal_content: Optional[dict] = None
     ignore_eos: bool = False
     language: Optional[str] = None
-    debug: bool = False
 
 
 @dataclass
 class RequestFuncOutput:
     """Output for requesting LLMs via API"""
-
-    no: int = 0
     generated_text: str = ""
     reasoning_content: str = ""
     success: bool = False
@@ -68,7 +64,7 @@ class RequestFuncOutput:
     itl: list = field(default_factory=list)  # list of inter-token latencies
     tpot: float = 0.0  # avg next-token latencies
     prompt_len: int = 0
-    prompt_tokens: int = 0  # 推理侧返回输入token数
+    prompt_tokens: int = 0 # 推理侧返回输入token数
     error: str = ""
 
 
@@ -78,19 +74,22 @@ async def async_request_eb_openai_chat_completions(
 ) -> RequestFuncOutput:
     """Request an LLM using EB OpenAI"""
     api_url = request_func_input.api_url
-    assert api_url.endswith(("completions", "profile")), "OpenAI Chat Completions API URL must end with 'completions'."
+    assert api_url.endswith(
+        ("completions", "profile")
+    ), "OpenAI Chat Completions API URL must end with 'completions'."
 
-    async with aiohttp.ClientSession(trust_env=True, timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
         content = [{"type": "text", "text": request_func_input.prompt}]
         if request_func_input.multi_modal_content:
             content.append(request_func_input.multi_modal_content)
         payload = {
-            "model": request_func_input.model,
+            "model": "default",
             "messages": request_func_input.history_QA,
             "stream": True,
             "stream_options": {
                 "include_usage": True,
-                "continuous_usage_stats": True,
+                "continuous_usage_stats": True
             },
         }
         # 超参由yaml传入
@@ -98,10 +97,6 @@ async def async_request_eb_openai_chat_completions(
 
         if request_func_input.ignore_eos:
             payload["ignore_eos"] = request_func_input.ignore_eos
-
-        if request_func_input.debug:
-            print(f"payload:{json.dumps(payload, ensure_ascii=False)}")
-
         headers = {
             "Content-Type": "application/json",
             "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
@@ -109,20 +104,21 @@ async def async_request_eb_openai_chat_completions(
 
         output = RequestFuncOutput()
         output.prompt_len = 0
-        output.no = request_func_input.no
 
         ttft = 0.0
         st = time.perf_counter()
         most_recent_timestamp = st
         try:
-            async with session.post(url=api_url, json=payload, headers=headers) as response:
+            async with session.post(url=api_url, json=payload,
+                                    headers=headers) as response:
                 if response.status == 200:
                     async for chunk_bytes in response.content:
                         chunk_bytes = chunk_bytes.strip()
                         if not chunk_bytes:
                             continue
 
-                        chunk = chunk_bytes.decode("utf-8").removeprefix("data: ")
+                        chunk = chunk_bytes.decode("utf-8").removeprefix(
+                            "data: ")
                         if chunk != "[DONE]":
                             # print("####chunk:", chunk, type(chunk))
                             timestamp = time.perf_counter()
@@ -136,20 +132,21 @@ async def async_request_eb_openai_chat_completions(
                                     ttft = timestamp - st
                                     output.ttft = ttft
                                     # cached_tokens
-                                    output.prompt_len = (
-                                        data["usage"].get("prompt_tokens_details", {}).get("cached_tokens", 0)
-                                    )
+                                    output.prompt_len = data["usage"]["prompt_tokens_details"]["cached_tokens"]
 
                                 # Decoding phase
                                 else:
-                                    output.itl.append(timestamp - most_recent_timestamp)
+                                    output.itl.append(timestamp -
+                                                      most_recent_timestamp)
 
                                 output.generated_text += content or ""
                                 output.reasoning_content += reason_content or ""
-                                output.arrival_time.append(choices[0].get("arrival_time", timestamp))
-                            elif usage := data.get("usage", {}):
-                                output.output_tokens = usage.get("completion_tokens", 0)
-                                output.prompt_tokens = usage.get("prompt_tokens", 0)
+                                output.arrival_time.append(choices[0].get("arrival_time"))
+                            elif usage := data.get("usage"):
+                                output.output_tokens = usage.get(
+                                    "completion_tokens")
+                                output.prompt_tokens = usage.get(
+                                    "prompt_tokens")
 
                             most_recent_timestamp = timestamp
 
@@ -162,12 +159,7 @@ async def async_request_eb_openai_chat_completions(
                     output.latency = most_recent_timestamp - st
                 else:
                     error_text = await response.text()
-                    print(
-                        "####error response:",
-                        error_text,
-                        "####payload:",
-                        payload,
-                    )
+                    print("####error response:", error_text, "####payload:", payload)
                     output.error = error_text or ""
                     output.success = False
         except Exception:
@@ -181,8 +173,6 @@ async def async_request_eb_openai_chat_completions(
                 f.write(str(output) + "\n")
     if pbar:
         pbar.update(1)
-    if request_func_input.debug:
-        print("#####final_output:", output)
     return output
 
 
@@ -196,14 +186,15 @@ async def async_request_eb_openai_completions(
         ("completions", "profile")
     ), "OpenAI Completions API URL must end with 'completions' or 'profile'."
 
-    async with aiohttp.ClientSession(trust_env=True, timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
         payload = {
-            "model": request_func_input.model,
+            "model": "default",
             "prompt": request_func_input.prompt,
             "stream": True,
             "stream_options": {
                 "include_usage": True,
-                "continuous_usage_stats": True,
+                "continuous_usage_stats": True
             },
         }
         # 超参由yaml传入
@@ -211,25 +202,19 @@ async def async_request_eb_openai_completions(
 
         if request_func_input.ignore_eos:
             payload["ignore_eos"] = request_func_input.ignore_eos
-
-        if request_func_input.debug:
-            print("payload:", json.dumps(payload, ensure_ascii=False))
-
         headers = {
-            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
-            "Content-Type": "application/json",
+            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"
         }
 
         output = RequestFuncOutput()
         output.prompt_len = request_func_input.prompt_len
-        output.no = request_func_input.no
 
         generated_text = ""
-        ttft = 0.0
         st = time.perf_counter()
         most_recent_timestamp = st
         try:
-            async with session.post(url=api_url, json=payload, headers=headers) as response:
+            async with session.post(url=api_url, json=payload,
+                                    headers=headers) as response:
                 if response.status == 200:
                     first_chunk_received = False
                     async for chunk_bytes in response.content:
@@ -237,10 +222,10 @@ async def async_request_eb_openai_completions(
                         if not chunk_bytes:
                             continue
 
-                        chunk = chunk_bytes.decode("utf-8").removeprefix("data: ")
+                        chunk = chunk_bytes.decode("utf-8").removeprefix(
+                            "data: ")
                         if chunk != "[DONE]":
                             # print("####chunk:", chunk, chunk.usage)
-                            timestamp = time.perf_counter()
                             data = json.loads(chunk)
 
                             # NOTE: Some completion API might have a last
@@ -250,40 +235,35 @@ async def async_request_eb_openai_completions(
                                 # Note that text could be empty here
                                 # e.g. for special tokens
                                 text = choices[0].get("text")
-
+                                timestamp = time.perf_counter()
                                 # First token
                                 if not first_chunk_received:
                                     first_chunk_received = True
-                                    ttft = timestamp - st
+                                    ttft = time.perf_counter() - st
                                     output.ttft = ttft
 
                                 # Decoding phase
                                 else:
-                                    output.itl.append(timestamp - most_recent_timestamp)
-
-                                generated_text += text or ""
+                                    output.itl.append(timestamp -
+                                                      most_recent_timestamp)
 
                                 most_recent_timestamp = timestamp
-                                output.arrival_time.append(choices[0].get("arrival_time", timestamp))
+                                output.arrival_time.append(choices[0].get("arrival_time"))
+                                generated_text += text or ""
                             elif usage := data.get("usage"):
-                                output.prompt_tokens = usage.get("prompt_tokens")
-                                output.output_tokens = usage.get("completion_tokens")
+                                output.prompt_tokens = usage.get(
+                                    "prompt_tokens")
+                                output.output_tokens = usage.get(
+                                    "completion_tokens")
                     if first_chunk_received:
                         output.success = True
                     else:
                         output.success = False
                         output.error = (
-                            "Never received a valid chunk to calculate TTFT." "This response will be marked as failed!"
-                        )
-
+                            "Never received a valid chunk to calculate TTFT."
+                            "This response will be marked as failed!")
                     output.generated_text = generated_text
                     output.latency = most_recent_timestamp - st
-
-                    if output.generated_text == "":
-                        output.success = False
-                        output.error = "No generated text found!"
-                    else:
-                        output.success = True
                 else:
                     output.error = response.reason or ""
                     output.success = False
@@ -292,9 +272,6 @@ async def async_request_eb_openai_completions(
             exc_info = sys.exc_info()
             output.error = "".join(traceback.format_exception(*exc_info))
 
-        if request_func_input.debug:
-            print(f"final_output:{output}")
-
     if pbar:
         pbar.update(1)
     return output
@@ -308,7 +285,8 @@ async def async_request_tgi(
     api_url = request_func_input.api_url
     assert api_url.endswith("generate_stream")
 
-    async with aiohttp.ClientSession(trust_env=True, timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
         params = {
             "max_new_tokens": request_func_input.output_len,
             "do_sample": True,
@@ -355,7 +333,8 @@ async def async_request_tgi(
 
                         # Decoding phase
                         else:
-                            output.itl.append(timestamp - most_recent_timestamp)
+                            output.itl.append(timestamp -
+                                              most_recent_timestamp)
 
                         most_recent_timestamp = timestamp
                         output.arrival_time.append(data["arrival_time"])
@@ -384,7 +363,8 @@ async def async_request_trt_llm(
     api_url = request_func_input.api_url
     assert api_url.endswith("generate_stream")
 
-    async with aiohttp.ClientSession(trust_env=True, timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
         payload = {
             "accumulate_tokens": True,
             "text_input": request_func_input.prompt,
@@ -409,7 +389,8 @@ async def async_request_trt_llm(
                         if not chunk_bytes:
                             continue
 
-                        chunk = chunk_bytes.decode("utf-8").removeprefix("data:")
+                        chunk = chunk_bytes.decode("utf-8").removeprefix(
+                            "data:")
 
                         data = json.loads(chunk)
                         output.generated_text += data["text_output"]
@@ -421,7 +402,8 @@ async def async_request_trt_llm(
 
                         # Decoding phase
                         else:
-                            output.itl.append(timestamp - most_recent_timestamp)
+                            output.itl.append(timestamp -
+                                              most_recent_timestamp)
 
                         most_recent_timestamp = timestamp
 
@@ -446,7 +428,8 @@ async def async_request_deepspeed_mii(
     pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
     """Request an LLM using Deepspeed MII"""
-    async with aiohttp.ClientSession(trust_env=True, timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
 
         payload = {
             "prompt": request_func_input.prompt,
@@ -464,16 +447,19 @@ async def async_request_deepspeed_mii(
 
         st = time.perf_counter()
         try:
-            async with session.post(url=request_func_input.api_url, json=payload) as response:
+            async with session.post(url=request_func_input.api_url,
+                                    json=payload) as response:
                 if response.status == 200:
                     parsed_resp = await response.json()
                     output.latency = time.perf_counter() - st
                     if "choices" in parsed_resp:
-                        output.generated_text = parsed_resp["choices"][0]["text"]
+                        output.generated_text = parsed_resp["choices"][0][
+                            "text"]
                     elif "text" in parsed_resp:
                         output.generated_text = parsed_resp["text"][0]
                     else:
-                        output.error = "Unexpected response format: " "neither 'choices' nor 'text' found"
+                        output.error = ("Unexpected response format: "
+                                        "neither 'choices' nor 'text' found")
                         output.success = False
                     output.success = True
                 else:
@@ -499,22 +485,26 @@ async def async_request_openai_completions(
         ("completions", "profile")
     ), "OpenAI Completions API URL must end with 'completions' or 'profile'."
 
-    async with aiohttp.ClientSession(trust_env=True, timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
         payload = {
-            "model": (request_func_input.model_name if request_func_input.model_name else request_func_input.model),
+            "model": request_func_input.model_name \
+                if request_func_input.model_name else request_func_input.model,
             "prompt": request_func_input.prompt,
             # "temperature": 0.0,
             "max_tokens": request_func_input.output_len,
             "logprobs": request_func_input.logprobs,
             "stream": True,
-            # "stream_options": {
+            #"stream_options": {
             #    "include_usage": True,
-            # },
+            #},
         }
         if request_func_input.ignore_eos:
             payload["ignore_eos"] = request_func_input.ignore_eos
 
-        headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"}
+        headers = {
+            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"
+        }
 
         output = RequestFuncOutput()
         output.prompt_len = request_func_input.prompt_len
@@ -523,7 +513,8 @@ async def async_request_openai_completions(
         st = time.perf_counter()
         most_recent_timestamp = st
         try:
-            async with session.post(url=api_url, json=payload, headers=headers) as response:
+            async with session.post(url=api_url, json=payload,
+                                    headers=headers) as response:
                 if response.status == 200:
                     first_chunk_received = False
                     async for chunk_bytes in response.content:
@@ -531,7 +522,8 @@ async def async_request_openai_completions(
                         if not chunk_bytes:
                             continue
 
-                        chunk = chunk_bytes.decode("utf-8").removeprefix("data: ")
+                        chunk = chunk_bytes.decode("utf-8").removeprefix(
+                            "data: ")
                         if chunk != "[DONE]":
                             # print("####chunk:", chunk, type(chunk))
                             data = json.loads(chunk)
@@ -552,19 +544,21 @@ async def async_request_openai_completions(
 
                                 # Decoding phase
                                 else:
-                                    output.itl.append(timestamp - most_recent_timestamp)
+                                    output.itl.append(timestamp -
+                                                      most_recent_timestamp)
 
                                 most_recent_timestamp = timestamp
                                 generated_text += text or ""
                             elif usage := data.get("usage"):
-                                output.output_tokens = usage.get("completion_tokens")
+                                output.output_tokens = usage.get(
+                                    "completion_tokens")
                     if first_chunk_received:
                         output.success = True
                     else:
                         output.success = False
                         output.error = (
-                            "Never received a valid chunk to calculate TTFT." "This response will be marked as failed!"
-                        )
+                            "Never received a valid chunk to calculate TTFT."
+                            "This response will be marked as failed!")
                     output.generated_text = generated_text
                     output.latency = most_recent_timestamp - st
                 else:
@@ -587,24 +581,25 @@ async def async_request_openai_audio(
     """Request an LLM using OpenAI"""
     # Lazy import without PlaceholderModule to avoid vllm dep.
     import soundfile
-
     api_url = request_func_input.api_url
     assert api_url.endswith(
-        ("transcriptions", "translations")
-    ), "OpenAI Chat Completions API URL must end with 'transcriptions' "
+        ("transcriptions", "translations"
+         )), "OpenAI Chat Completions API URL must end with 'transcriptions' "
     "or `translations`."
 
-    async with aiohttp.ClientSession(trust_env=True, timeout=AIOHTTP_TIMEOUT) as session:
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
         content = [{"type": "text", "text": request_func_input.prompt}]
         payload = {
-            "model": (request_func_input.model_name if request_func_input.model_name else request_func_input.model),
+            "model": request_func_input.model_name \
+                if request_func_input.model_name else request_func_input.model,
             "temperature": 0.0,
             "max_completion_tokens": request_func_input.output_len,
             "stream": True,
             "language": "en",
             # Flattened due to multipart/form-data
             "stream_include_usage": True,
-            "stream_continuous_usage_stats": True,
+            "stream_continuous_usage_stats": True
         }
         if request_func_input.extra_body:
             payload.update(request_func_input.extra_body)
@@ -619,9 +614,9 @@ async def async_request_openai_audio(
             buffer.seek(0)
             return buffer
 
-        with to_bytes(*request_func_input.multi_modal_content["audio"]) as f:
+        with to_bytes(*request_func_input.multi_modal_content['audio']) as f:
             form = aiohttp.FormData()
-            form.add_field("file", f, content_type="audio/wav")
+            form.add_field('file', f, content_type='audio/wav')
             for key, value in payload.items():
                 form.add_field(key, str(value))
 
@@ -633,20 +628,24 @@ async def async_request_openai_audio(
             st = time.perf_counter()
             most_recent_timestamp = st
             try:
-                async with session.post(url=api_url, data=form, headers=headers) as response:
+                async with session.post(url=api_url,
+                                        data=form,
+                                        headers=headers) as response:
                     if response.status == 200:
                         async for chunk_bytes in response.content:
                             chunk_bytes = chunk_bytes.strip()
                             if not chunk_bytes:
                                 continue
 
-                            chunk = chunk_bytes.decode("utf-8").removeprefix("data: ")
+                            chunk = chunk_bytes.decode("utf-8").removeprefix(
+                                "data: ")
                             if chunk != "[DONE]":
                                 timestamp = time.perf_counter()
                                 data = json.loads(chunk)
 
                                 if choices := data.get("choices"):
-                                    content = choices[0]["delta"].get("content")
+                                    content = choices[0]["delta"].get(
+                                        "content")
                                     # First token
                                     if ttft == 0.0:
                                         ttft = timestamp - st
@@ -654,11 +653,13 @@ async def async_request_openai_audio(
 
                                     # Decoding phase
                                     else:
-                                        output.itl.append(timestamp - most_recent_timestamp)
+                                        output.itl.append(
+                                            timestamp - most_recent_timestamp)
 
                                     generated_text += content or ""
                                 elif usage := data.get("usage"):
-                                    output.output_tokens = usage.get("completion_tokens")
+                                    output.output_tokens = usage.get(
+                                        "completion_tokens")
 
                                 most_recent_timestamp = timestamp
 
@@ -692,11 +693,8 @@ ASYNC_REQUEST_FUNCS = {
 }
 
 OPENAI_COMPATIBLE_BACKENDS = [
-    k
-    for k, v in ASYNC_REQUEST_FUNCS.items()
-    if v
-    in (
-        async_request_openai_completions,
-        async_request_eb_openai_chat_completions,
-    )
+    k for k, v in ASYNC_REQUEST_FUNCS.items()
+    if v in (async_request_openai_completions,
+             async_request_eb_openai_chat_completions)
 ]
+

benchmarks/benchmark_dataset.py

@@ -26,10 +26,10 @@ from abc import ABC, abstractmethod
 from collections.abc import Mapping
 from dataclasses import dataclass
 from io import BytesIO
-from typing import Any, Optional, Union
-
+from typing import Any, Callable, Optional, Union
 from PIL import Image
 
+
 logger = logging.getLogger(__name__)
 
 
@@ -39,7 +39,6 @@ class SampleRequest:
     Represents a single inference request for benchmarking.
     """
 
-    no: int
     prompt: Union[str, Any]
     history_QA: Union[str, Any]
     json_data: Optional[dict]
@@ -49,7 +48,6 @@ class SampleRequest:
 
 class BenchmarkDataset(ABC):
     """BenchmarkDataset"""
-
     DEFAULT_SEED = 0
     IS_MULTIMODAL = False
 
@@ -57,7 +55,6 @@ class BenchmarkDataset(ABC):
         self,
         dataset_path: Optional[str] = None,
         random_seed: int = DEFAULT_SEED,
-        shuffle: bool = False,
         hyperparameter_path: Optional[str] = None,
     ) -> None:
         """
@@ -71,9 +68,9 @@ class BenchmarkDataset(ABC):
         self.dataset_path = dataset_path
         # Set the random seed, ensuring that a None value is replaced with the
         # default seed.
-        self.random_seed = random_seed if random_seed is not None else self.DEFAULT_SEED
+        self.random_seed = (random_seed
+                            if random_seed is not None else self.DEFAULT_SEED)
         self.data = None
-        self.shuffle = shuffle
         self.hyperparameter_path = hyperparameter_path
         self.hyperparameters = {}
 
@@ -88,7 +85,8 @@ class BenchmarkDataset(ABC):
             NotImplementedError: If a subclass does not implement this method.
         """
         # TODO (jenniferzhao): add support for downloading data
-        raise NotImplementedError("load_data must be implemented in subclasses.")
+        raise NotImplementedError(
+            "load_data must be implemented in subclasses.")
 
     @abstractmethod
     def sample(self, num_requests: int) -> list[SampleRequest]:
@@ -107,7 +105,8 @@ class BenchmarkDataset(ABC):
         """
         raise NotImplementedError("sample must be implemented in subclasses.")
 
-    def maybe_oversample_requests(self, requests: list[SampleRequest], num_requests: int) -> None:
+    def maybe_oversample_requests(self, requests: list[SampleRequest],
+                                  num_requests: int) -> None:
         """
         Oversamples the list of requests if its size is less than the desired
         number.
@@ -118,9 +117,11 @@ class BenchmarkDataset(ABC):
         """
         if len(requests) < num_requests:
             random.seed(self.random_seed)
-            additional = random.choices(requests, k=num_requests - len(requests))
+            additional = random.choices(requests,
+                                        k=num_requests - len(requests))
             requests.extend(additional)
-            logger.info("Oversampled requests to reach %d total samples.", num_requests)
+            logger.info("Oversampled requests to reach %d total samples.",
+                        num_requests)
 
 
 def is_valid_sequence(
@@ -140,12 +141,14 @@ def is_valid_sequence(
     """
     # Check for invalid conditions
     prompt_too_short = prompt_len < min_len
-    output_too_short = (not skip_min_output_len_check) and (output_len < min_len)
+    output_too_short = (not skip_min_output_len_check) and (output_len
+                                                            < min_len)
     prompt_too_long = prompt_len > max_prompt_len
     combined_too_long = (prompt_len + output_len) > max_total_len
 
     # Return True if none of the invalid conditions are met
-    return not (prompt_too_short or output_too_short or prompt_too_long or combined_too_long)
+    return not (prompt_too_short or output_too_short or prompt_too_long
+                or combined_too_long)
 
 
 def process_image(image: Any) -> Mapping[str, Any]:
@@ -168,25 +171,28 @@ def process_image(image: Any) -> Mapping[str, Any]:
     Raises:
         ValueError: If the input is not a supported type.
     """
-    if isinstance(image, dict) and "bytes" in image:
-        image = Image.open(BytesIO(image["bytes"]))
+    if isinstance(image, dict) and 'bytes' in image:
+        image = Image.open(BytesIO(image['bytes']))
     if isinstance(image, Image.Image):
         image = image.convert("RGB")
         with io.BytesIO() as image_data:
             image.save(image_data, format="JPEG")
-            image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8")
+            image_base64 = base64.b64encode(
+                image_data.getvalue()).decode("utf-8")
         return {
             "type": "image_url",
-            "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"},
+            "image_url": {
+                "url": f"data:image/jpeg;base64,{image_base64}"
+            },
         }
 
     if isinstance(image, str):
-        image_url = image if image.startswith(("http://", "file://")) else f"file://{image}"
+        image_url = (image if image.startswith(
+            ("http://", "file://")) else f"file://{image}")
         return {"type": "image_url", "image_url": {"url": image_url}}
 
-    raise ValueError(
-        f"Invalid image input {image}. Must be a PIL.Image.Image" " or str or dictionary with raw image bytes."
-    )
+    raise ValueError(f"Invalid image input {image}. Must be a PIL.Image.Image"
+                     " or str or dictionary with raw image bytes.")
 
 
 class EBDataset(BenchmarkDataset):
@@ -213,10 +219,6 @@ class EBDataset(BenchmarkDataset):
         with open(self.dataset_path, encoding="utf-8") as f:
             self.data = [json.loads(i.strip()) for i in f.readlines()]
 
-        if self.shuffle:
-            random.seed(self.random_seed)
-            random.shuffle(self.data)
-
     def sample(
         self,
         num_requests: int,
@@ -227,7 +229,6 @@ class EBDataset(BenchmarkDataset):
         **kwargs,
     ) -> list:
         samples: list = []
-        cnt = 1
         for entry in self.data:
             if len(samples) >= num_requests:
                 break
@@ -241,17 +242,15 @@ class EBDataset(BenchmarkDataset):
             new_output_len = int(entry["max_dec_len"])
 
             if enable_multimodal_chat:
-                prompt = self.apply_multimodal_chat_transformation(prompt, None)
+                prompt = self.apply_multimodal_chat_transformation(
+                    prompt, None)
             samples.append(
                 SampleRequest(
-                    no=cnt,
                     prompt=prompt,
                     prompt_len=self.prompt_len,
                     history_QA=[],
                     expected_output_len=new_output_len,
-                )
-            )
-            cnt += 1
+                ))
 
         self.maybe_oversample_requests(samples, num_requests)
         return samples
@@ -262,7 +261,6 @@ class EBChatDataset(BenchmarkDataset):
     Implements the ShareGPT dataset.  Loads data from a JSON file and generates
     sample requests based on conversation turns.
     """
-
     prompt_len: int
 
     def __init__(self, **kwargs) -> None:
@@ -276,10 +274,6 @@ class EBChatDataset(BenchmarkDataset):
         with open(self.dataset_path, encoding="utf-8") as f:
             self.data = [json.loads(i.strip()) for i in f.readlines()]
 
-        if self.shuffle:
-            random.seed(self.random_seed)
-            random.shuffle(self.data)
-
     def sample(
         self,
         num_requests: int,
@@ -290,7 +284,6 @@ class EBChatDataset(BenchmarkDataset):
         **kwargs,
     ) -> list:
         samples: list = []
-        cnt = 1
         for entry in self.data:
             if len(samples) >= num_requests:
                 break
@@ -300,18 +293,17 @@ class EBChatDataset(BenchmarkDataset):
             new_output_len = int(entry.get("max_tokens", 12288))
 
             if enable_multimodal_chat:
-                prompt = self.apply_multimodal_chat_transformation(prompt, None)
+                prompt = self.apply_multimodal_chat_transformation(
+                    prompt, None)
             samples.append(
                 SampleRequest(
-                    no=cnt,
                     json_data=json_data,
                     prompt=prompt,
                     prompt_len=0,
                     history_QA=history_QA,
                     expected_output_len=new_output_len,
-                )
-            )
-            cnt += 1
+                ))
 
         self.maybe_oversample_requests(samples, num_requests)
         return samples
+

benchmarks/benchmark_mtp.py

@@ -1,178 +0,0 @@
-"""
-# Copyright (c) 2025  PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License"
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""
-
-import argparse
-import asyncio
-import contextlib
-import os
-from typing import Union
-
-from benchmark_dataset import EBChatDataset, EBDataset
-from benchmark_serving import benchmark
-
-
-def prepare_input_requests(num_prompts: int, dataset_name: str, dataset_path: str) -> Union[EBDataset, EBChatDataset]:
-    dataset_mapping = {
-        "EB": lambda: EBDataset(dataset_path=dataset_path).sample(num_requests=num_prompts),
-        "EBChat": lambda: EBChatDataset(dataset_path=dataset_path).sample(num_requests=num_prompts),
-    }
-
-    try:
-        input_requests = dataset_mapping[dataset_name]()
-    except KeyError as err:
-        raise ValueError(f"Unknown dataset: {dataset_name}") from err
-
-    return input_requests
-
-
-class FakeTokenizer:
-    def encode(self, text: str, add_special_tokens: bool = False):
-        return []
-
-
-def send_one_batch(base_url, max_concurrency, input_requests, disable_tqdm):
-    selected_percentile_metrics = ["s_itl"]
-    selected_percentiles = []
-    # Run benchmark
-    results = asyncio.run(
-        benchmark(
-            backend="openai-chat",
-            api_url=f"{base_url}/v1/chat/completions",
-            base_url=base_url,
-            model_id="default",
-            model_name="default",
-            input_requests=input_requests,
-            hyper_parameters={},
-            logprobs=None,
-            request_rate=float("inf"),
-            burstiness=1.0,
-            disable_tqdm=disable_tqdm,
-            profile=False,
-            selected_percentile_metrics=selected_percentile_metrics,
-            selected_percentiles=selected_percentiles,
-            ignore_eos=False,
-            goodput_config_dict=None,
-            max_concurrency=max_concurrency,
-            lora_modules=None,
-            extra_body=None,
-        )
-    )
-
-    record = {
-        "mean_s_itl_ms": results["mean_s_itl_ms"],
-    }
-
-    return record
-
-
-def calculate_speedup(acceptance_rate, draft_token_step, t_ori, t_mtp):
-
-    tmp = 0.0
-    for i in range(draft_token_step):
-        tmp += pow(acceptance_rate, i + 1)
-
-    r_ac = tmp / (1 + tmp)
-
-    return t_ori / ((1 - r_ac) * t_mtp)
-
-
-def main(args):
-    base_url = f"http://{args.host}:{args.port}"
-
-    input_requests = prepare_input_requests(args.num_prompts, args.dataset_name, args.dataset_path)
-
-    if len(args.max_concurrency) != len(args.s_itl_base_model):
-        raise ValueError("--max_concurrency should be same length as --s_itl_base_model")
-
-    for max_concurrency, s_itl in zip(args.max_concurrency, args.s_itl_base_model):
-        # Wramup
-        print("Starting warmup...")
-        with open(os.devnull, "w") as f:
-            with contextlib.redirect_stdout(f):
-                send_one_batch(
-                    base_url,
-                    max_concurrency,
-                    input_requests[0:max_concurrency],
-                    True,
-                )
-
-        # Benchmark
-        record = send_one_batch(base_url, max_concurrency, input_requests, False)
-
-        metric_header = "Speed up"
-        print("{s:{c}^{n}}".format(s=metric_header, n=50, c="-"))
-        for draft_token_step in args.draft_token_steps:
-            speedup = calculate_speedup(
-                args.acceptance_rate,
-                draft_token_step,
-                s_itl,
-                record["mean_s_itl_ms"],
-            )
-            print("{:<40} {:<10.2f}".format(f"Speed up on {draft_token_step} steps draft", speedup))
-        print("=" * 50)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--host",
-        type=str,
-        default="127.0.0.1",
-    )
-    parser.add_argument(
-        "--port",
-        type=str,
-        default="8000",
-    )
-    parser.add_argument(
-        "--max-concurrency",
-        type=int,
-        nargs="+",
-        default=(1, 2, 4, 8, 16, 32),
-    )
-    parser.add_argument(
-        "--num-prompts",
-        type=int,
-        default=128,
-    )
-    parser.add_argument(
-        "--acceptance-rate",
-        type=float,
-        default=0.8,
-    )
-    parser.add_argument(
-        "--draft-token-steps",
-        type=int,
-        nargs="+",
-        default=(1, 2),
-    )
-    parser.add_argument(
-        "--s_itl-base-model",
-        type=float,
-        nargs="+",
-    )
-    parser.add_argument(
-        "--dataset-name",
-        type=str,
-        default="EBChat",
-    )
-    parser.add_argument(
-        "--dataset-path",
-        type=str,
-    )
-    args = parser.parse_args()
-
-    main(args)

benchmarks/benchmark_utils.py

@@ -24,11 +24,9 @@ import os
 from typing import Any
 
 
-def convert_to_pytorch_benchmark_format(
-    args: argparse.Namespace,
-    metrics: dict[str, list],
-    extra_info: dict[str, Any],
-) -> list:
+def convert_to_pytorch_benchmark_format(args: argparse.Namespace,
+                                        metrics: dict[str, list],
+                                        extra_info: dict[str, Any]) -> list:
     """
     Save the benchmark results in the format used by PyTorch OSS benchmark with
     on metric per record
@@ -56,10 +54,12 @@ def convert_to_pytorch_benchmark_format(
             },
         }
 
-        tp = record["benchmark"]["extra_info"]["args"].get("tensor_parallel_size")
+        tp = record["benchmark"]["extra_info"]["args"].get(
+            "tensor_parallel_size")
         # Save tensor_parallel_size parameter if it's part of the metadata
         if not tp and "tensor_parallel_size" in extra_info:
-            record["benchmark"]["extra_info"]["args"]["tensor_parallel_size"] = extra_info["tensor_parallel_size"]
+            record["benchmark"]["extra_info"]["args"][
+                "tensor_parallel_size"] = extra_info["tensor_parallel_size"]
 
         records.append(record)
 
@@ -68,7 +68,6 @@ def convert_to_pytorch_benchmark_format(
 
 class InfEncoder(json.JSONEncoder):
     """InfEncoder"""
-
     def clear_inf(self, o: Any):
         """clear_inf"""
         if isinstance(o, dict):
@@ -88,3 +87,4 @@ def write_to_json(filename: str, records: list) -> None:
     """write_to_json"""
     with open(filename, "w") as f:
         json.dump(records, f, cls=InfEncoder)
+

benchmarks/requirements.txt

@@ -3,4 +3,3 @@ tqdm
 numpy
 Pillow
 pyyaml
-requests

benchmarks/yaml/eb45-21B-vl-128k-wint4-h800-tp1.yaml

@@ -7,4 +7,4 @@ tensor_parallel_size: 1
 enable_chunked_prefill: True
 max_num_batched_tokens: 384
 quantization: wint4
-reasoning_parser: ernie-45-vl
+reasoning_parser: ernie-45-vl

benchmarks/yaml/eb45-32k-w4a8c8-tp4_decode.yaml

@@ -12,4 +12,4 @@ rdma_comm_ports: "7671,7672,7673,7674"
 pd_comm_port: "2334"
 max_num_batched_tokens: 384
 max_num_partial_prefills: 3
-max_long_partial_prefills: 3
+max_long_partial_prefills: 3

benchmarks/yaml/eb45-32k-w4a8c8-tp4_prefill.yaml

@@ -9,4 +9,4 @@ cache_queue_port: 55664
 engine_worker_queue_port: 6677
 cache_transfer_protocol: "rdma,ipc"
 rdma_comm_ports: "7675,7676,7677,7678"
-pd_comm_port: "2333"
+pd_comm_port: "2333"

benchmarks/yaml/eb45-32k-wint4-a800-tp4.yaml

@@ -3,4 +3,3 @@ max_num_seqs: 96
 gpu_memory_utilization: 0.9
 kv_cache_ratio: 0.71
 tensor_parallel_size: 4
-quantization: wint4

benchmarks/yaml/eb45-32k-wint4-h800-dp8_prefill.yaml

@@ -10,4 +10,4 @@ engine_worker_queue_port: 6677
 num_gpu_blocks_override: 1024
 cache_transfer_protocol: "rdma"
 rdma_comm_ports: "7671,7672,7673,7674,7675,7676,7677,7678"
-pd_comm_port: "2334"
+pd_comm_port: "2334"

benchmarks/yaml/eb45-32k-wint4-mtp-tp4-decode.yaml

@@ -10,4 +10,4 @@ splitwise_role: decode
 engine_worker_queue_port: 6678
 cache_transfer_protocol: "rdma,ipc"
 rdma_comm_ports: "7671,7672,7673,7674"
-pd_comm_port: "2334"
+pd_comm_port: "2334"

benchmarks/yaml/eb45-32k-wint4-mtp-tp4-prefill.yaml

@@ -9,4 +9,4 @@ cache_queue_port: 55664
 engine_worker_queue_port: 6677
 cache_transfer_protocol: "rdma,ipc"
 rdma_comm_ports: "7675,7676,7677,7678"
-pd_comm_port: "2333"
+pd_comm_port: "2333"

benchmarks/yaml/eb45-32k-wint4-tp4_decode.yaml

@@ -12,4 +12,4 @@ rdma_comm_ports: "7671,7672,7673,7674"
 pd_comm_port: "2334"
 max_num_batched_tokens: 384
 max_num_partial_prefills: 3
-max_long_partial_prefills: 3
+max_long_partial_prefills: 3

benchmarks/yaml/eb45-32k-wint4-tp4_prefill.yaml

@@ -9,4 +9,4 @@ cache_queue_port: 55664
 engine_worker_queue_port: 6677
 cache_transfer_protocol: "rdma,ipc"
 rdma_comm_ports: "7675,7676,7677,7678"
-pd_comm_port: "2333"
+pd_comm_port: "2333"

benchmarks/yaml/eb45-32k-wint8-a800-tp8.yaml

@@ -3,4 +3,3 @@ max_num_seqs: 96
 gpu_memory_utilization: 0.9
 kv_cache_ratio: 0.71
 tensor_parallel_size: 8
-quantization: wint8

benchmarks/yaml/eb45-8k-fp8-tp1-dp8_ep.yaml

@@ -1,6 +0,0 @@
-num_gpu_blocks_override: 1024
-max_model_len: 8192
-max_num_seqs: 64
-data_parallel_size: 8
-tensor_parallel_size: 1
-enable_expert_parallel: True

benchmarks/yaml/eb45t_0dot3b-32k-bf16-a30-tp1-static.yaml

@@ -2,5 +2,4 @@ max_model_len: 32768
 max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/eb45t_0dot3b-32k-bf16-h800-tp1-static.yaml

@@ -2,5 +2,4 @@ max_model_len: 32768
 max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/eb45t_0dot3b-32k-wint8-a30-tp1-static.yaml

@@ -3,5 +3,4 @@ max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
 quantization: wint8
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/eb45t_0dot3b-32k-wint8-h800-tp1-static.yaml

@@ -3,5 +3,4 @@ max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
 quantization: wint8
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/eb45t_21b-32k-bf16-h800-tp1-static.yaml

@@ -2,5 +2,4 @@ max_model_len: 32768
 max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/eb45t_21b-32k-wint4-h800-tp1-static.yaml

@@ -3,5 +3,4 @@ max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
 quantization: wint4
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/eb45t_300b-32k-wint4-h800-tp4-static.yaml

@@ -3,5 +3,4 @@ max_num_seqs: 96
 gpu_memory_utilization: 0.9
 kv_cache_ratio: 0.71
 tensor_parallel_size: 4
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/qwen2_7b-32k-bf16-a30-tp1-static.yaml

@@ -2,5 +2,4 @@ max_model_len: 32768
 max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/qwen2_7b-32k-bf16-h800-tp1-static.yaml

@@ -2,5 +2,4 @@ max_model_len: 32768
 max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/qwen2_7b-32k-fp8-h800-tp1-static.yaml

@@ -3,5 +3,4 @@ max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
 quantization: wfp8afp8
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/qwen3_0dot6b-32k-bf16-a30-tp1-static.yaml

@@ -2,5 +2,4 @@ max_model_len: 32768
 max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/qwen3_0dot6b-32k-bf16-h800-tp1-static.yaml

@@ -2,5 +2,4 @@ max_model_len: 32768
 max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/qwen3_0dot6b-32k-wint8-a30-tp1-static.yaml

@@ -3,5 +3,4 @@ max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
 quantization: wint8
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/qwen3_0dot6b-32k-wint8-h800-tp1-static.yaml

@@ -3,5 +3,4 @@ max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
 quantization: wint8
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/qwen3_30b-32k-bf16-h800-tp1-static.yaml

@@ -2,5 +2,4 @@ max_model_len: 32768
 max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/qwen3_30b-32k-wint4-h800-tp1-static.yaml

@@ -3,5 +3,4 @@ max_num_seqs: 128
 kv_cache_ratio: 0.75
 tensor_parallel_size: 1
 quantization: wint4
-graph_optimization_config:
-  graph_opt_level: 1
+enable_static_graph_inference: True

benchmarks/yaml/qwen3moe235b-32k-wint4-h800-tp4.yaml

@@ -3,4 +3,4 @@ max_num_seqs: 75
 gpu_memory_utilization: 0.85
 kv_cache_ratio: 0.75
 quantization: wint4
-tensor_parallel_size: 4
+tensor_parallel_size: 4

benchmarks/yaml/qwen3moe235b-32k-wint8-h800-tp4.yaml

@@ -3,4 +3,4 @@ max_num_seqs: 25
 gpu_memory_utilization: 0.9
 kv_cache_ratio: 0.75
 quantization: wint8
-tensor_parallel_size: 4
+tensor_parallel_size: 4

benchmarks/yaml/request_yaml/quick_benchmark.yaml

@@ -1,3 +0,0 @@
-metadata:
-  min_tokens: 32
-max_tokens: 33

benchmarks/yaml/request_yaml/qwen2-32k.yaml

@@ -5,4 +5,4 @@ metadata:
 max_tokens: 12288
 repetition_penalty: 1.05
 frequency_penalty: 0
-presence_penalty: 0
+presence_penalty: 0

benchmarks/yaml/request_yaml/qwen3-32k.yaml

@@ -5,4 +5,4 @@ metadata:
 max_tokens: 12288
 repetition_penalty: 1.0
 frequency_penalty: 0
-presence_penalty: 1.5
+presence_penalty: 1.5

benchmarks/yaml/request_yaml/vLLM_default.yaml

@@ -1,11 +0,0 @@
-top_p: 1.0
-temperature: 1.0
-metadata:
-  min_tokens: 1
-max_tokens: 30721
-repetition_penalty: 1.0
-frequency_penalty: 0
-presence_penalty: 0
-skip_special_tokens: false
-chat_template_kwargs:
-  enable_thinking: true

benchmarks/yaml/request_yaml/x1.yaml

@@ -1,8 +0,0 @@
-top_p: 0.95
-temperature: 0.6
-metadata:
-  min_tokens: 1
-max_tokens: 65535
-repetition_penalty: 1.0
-frequency_penalty: 0
-presence_penalty: 0

benchmarks/yaml/x1-32k-wint8-p800-tp8.yaml

@@ -3,4 +3,4 @@ max_num_seqs: 64
 gpu_memory_utilization: 0.9
 tensor_parallel_size: 8
 quantization: wint8
-reasoning_parser: ernie-x1
+reasoning_parser: ernie-x1

benchmarks/yaml/x1-64k-w4a8c8-tp4.yaml

@@ -1,10 +0,0 @@
-reasoning-parser: ernie_x1
-tool_call_parser: ernie_x1
-tensor_parallel_size: 4
-max_model_len: 65536
-max_num_seqs: 128
-enable_prefix_caching: True
-enable_chunked_prefill: True
-gpu_memory_utilization: 0.85
-use_cudagraph: True
-enable_custom_all_reduce: True

build.sh

@@ -18,9 +18,6 @@ BUILD_WHEEL=${1:-1}
 PYTHON_VERSION=${2:-"python"}
 export python=$PYTHON_VERSION
 FD_CPU_USE_BF16=${3:-"false"}
-# FD_BUILDING_ARCS: Specify target CUDA architectures for custom ops, e.g., "[80, 90, 100]".
-# For SM90 (Hopper), use 90. For SM100 (Blackwell), use 100.
-# These will be translated to 90a / 100a in setup_ops.py for specific features.
 FD_BUILDING_ARCS=${4:-""}
 
 
@@ -34,6 +31,7 @@ EGG_DIR="fastdeploy.egg-info"
 
 # custom_ops directory config
 OPS_SRC_DIR="custom_ops"
+OPS_TMP_DIR_BASE="tmp_base"
 OPS_TMP_DIR="tmp"
 
 # command line log config
@@ -70,6 +68,7 @@ function copy_ops(){
     PY_VERSION="py${PY_MAIN_VERSION}.${PY_SUB_VERSION}"
     SYSTEM_VERSION=`${python} -c "import platform; print(platform.system().lower())"`
     PROCESSOR_VERSION=`${python} -c "import platform; print(platform.processor())"`
+    WHEEL_BASE_NAME="fastdeploy_base_ops-${OPS_VERSION}-${PY_VERSION}-${SYSTEM_VERSION}-${PROCESSOR_VERSION}.egg"
     WHEEL_NAME="fastdeploy_ops-${OPS_VERSION}-${PY_VERSION}-${SYSTEM_VERSION}-${PROCESSOR_VERSION}.egg"
     WHEEL_CPU_NAME="fastdeploy_cpu_ops-${OPS_VERSION}-${PY_VERSION}-${SYSTEM_VERSION}-${PROCESSOR_VERSION}.egg"
     is_rocm=`$python -c "import paddle; print(paddle.is_compiled_with_rocm())"`
@@ -79,11 +78,13 @@ function copy_ops(){
       echo -e "ROCM ops have been copy to fastdeploy"
       return
     fi
+    mkdir -p ../fastdeploy/model_executor/ops/base
     is_cuda=`$python -c "import paddle; print(paddle.is_compiled_with_cuda())"`
     if [ "$is_cuda" = "True" ]; then
       DEVICE_TYPE="gpu"
+      cp -r ./${OPS_TMP_DIR_BASE}/${WHEEL_BASE_NAME}/* ../fastdeploy/model_executor/ops/base
       cp -r ./${OPS_TMP_DIR}/${WHEEL_NAME}/* ../fastdeploy/model_executor/ops/gpu
-      echo -e "CUDA ops have been copy to fastdeploy"
+      echo -e "BASE and CUDA ops have been copy to fastdeploy"
       return
     fi
 
@@ -103,42 +104,20 @@ function copy_ops(){
       return
     fi
 
-    if_corex=`$python -c "import paddle; print(paddle.is_compiled_with_custom_device(\"iluvatar_gpu\"))"`
-    if [ "$if_corex" = "True" ]; then
-      DEVICE_TYPE="iluvatar-gpu"
-      cp -r ./${OPS_TMP_DIR}/${WHEEL_NAME}/* ../fastdeploy/model_executor/ops/iluvatar
-      echo -e "Iluvatar ops have been copy to fastdeploy"
-      return
-    fi
-
-    is_gcu=`$python -c "import paddle; print(paddle.is_compiled_with_custom_device('gcu'))"`
-    if [ "$is_gcu" = "True" ]; then
-      DEVICE_TYPE="gcu"
-      cp -r ${OPS_TMP_DIR}/${WHEEL_NAME}/* ../fastdeploy/model_executor/ops/gcu
-      echo -e "gcu ops have been copy to fastdeploy"
-      return
-    fi
-
-    is_maca=`$python -c "import paddle; print(paddle.device.is_compiled_with_custom_device('metax_gpu'))"`
-    if [ "$is_maca" = "True" ]; then
-      DEVICE_TYPE="metax_gpu"
-      mkdir -p ../fastdeploy/model_executor/ops/base
-      cp -r ./${OPS_TMP_DIR_BASE}/${WHEEL_BASE_NAME}/* ../fastdeploy/model_executor/ops/base
-      cp -r ./${OPS_TMP_DIR}/${WHEEL_NAME}/* ../fastdeploy/model_executor/ops/gpu
-      echo -e "MACA ops have been copy to fastdeploy"
-      return
-    fi
-
     DEVICE_TYPE="cpu"
+    cp -r ./${OPS_TMP_DIR_BASE}/${WHEEL_BASE_NAME}/* ../fastdeploy/model_executor/ops/base
     cd ../../../../
     cp -r ${OPS_TMP_DIR}/${WHEEL_CPU_NAME}/* ../fastdeploy/model_executor/ops/cpu
-    echo -e "CPU ops have been copy to fastdeploy"
+    echo -e "BASE and CPU ops have been copy to fastdeploy"
     return
 }
 
 function build_and_install_ops() {
   cd $OPS_SRC_DIR
   export no_proxy=bcebos.com,paddlepaddle.org.cn,${no_proxy}
+  echo -e "${BLUE}[build]${NONE} build and install fastdeploy_base_ops..."
+  ${python} setup_ops_base.py install --install-lib ${OPS_TMP_DIR_BASE}
+  find ${OPS_TMP_DIR_BASE} -type f -name "*.o" -exec rm -f {} \;
   echo -e "${BLUE}[build]${NONE} build and install fastdeploy_ops..."
   TMP_DIR_REAL_PATH=`readlink -f ${OPS_TMP_DIR}`
   is_xpu=`$python -c "import paddle; print(paddle.is_compiled_with_xpu())"`
@@ -184,24 +163,17 @@ function build_and_install() {
     exit 1
   fi
   echo -e "${BLUE}[build]${NONE} ${GREEN}build fastdeploy wheel success${NONE}\n"
-}
 
-function version_info() {
-  output_file="fastdeploy/version.txt"
-  fastdeploy_git_commit_id=$(git rev-parse HEAD)
-  paddle_version=$(${python} -c "import paddle; print(paddle.__version__)")
-  paddle_git_commit_id=$(${python} -c "import paddle; print(paddle.__git_commit__)")
-  cuda_version="nvcc-not-installed"
-  if command -v nvcc &> /dev/null; then
-    cuda_version=$(nvcc -V | grep -Po "(?<=release )[\d.]+(?=, V)")
+  echo -e "${BLUE}[install]${NONE} installing fastdeploy..."
+  cd $DIST_DIR
+  find . -name "fastdeploy*.whl" | xargs ${python} -m pip install
+  if [ $? -ne 0 ]; then
+    cd ..
+    echo -e "${RED}[FAIL]${NONE} install fastdeploy wheel failed"
+    exit 1
   fi
-  cxx_version=$(g++ --version | head -n 1 | grep -Po "(?<=\) )[\d.]+")
-
-  echo "fastdeploy GIT COMMIT ID: $fastdeploy_git_commit_id" > $output_file
-  echo "Paddle version: $paddle_version" >> $output_file
-  echo "Paddle GIT COMMIT ID: $paddle_git_commit_id" >> $output_file
-  echo "CUDA version: $cuda_version" >> $output_file
-  echo "CXX compiler version: $cxx_version" >> $output_file
+  echo -e "${BLUE}[install]${NONE} ${GREEN}fastdeploy install success${NONE}\n"
+  cd ..
 }
 
 function cleanup() {
@@ -212,6 +184,7 @@ function cleanup() {
   fi
 
   rm -rf $OPS_SRC_DIR/$BUILD_DIR $OPS_SRC_DIR/$EGG_DIR
+  rm -rf $OPS_SRC_DIR/$OPS_TMP_DIR_BASE
   rm -rf $OPS_SRC_DIR/$OPS_TMP_DIR
 }
 
@@ -234,7 +207,6 @@ if [ "$BUILD_WHEEL" -eq 1 ]; then
   set -e
 
   init
-  version_info
   build_and_install_ops
   build_and_install
   cleanup
@@ -265,7 +237,6 @@ if [ "$BUILD_WHEEL" -eq 1 ]; then
 else
   init
   build_and_install_ops
-  version_info
   rm -rf $BUILD_DIR $EGG_DIR $DIST_DIR
   rm -rf $OPS_SRC_DIR/$BUILD_DIR $OPS_SRC_DIR/$EGG_DIR
 fi

custom_ops/0001-DeepGEMM-95e81b3.patch

@@ -26,7 +26,7 @@ index 15b22ca..63e7fb7 100644
 @@ -1,4 +1,4 @@
 -import torch
 +import paddle
-
+ 
  from . import jit
  from .jit_kernels import (
 diff --git a/deep_gemm/include/deep_gemm/scheduler.cuh b/deep_gemm/include/deep_gemm/scheduler.cuh
@@ -53,7 +53,7 @@ index c17d466..6fdc52f 100644
 -from torch.utils.cpp_extension import CUDA_HOME
 +from ..paddle_utils import CUDA_HOME
  from typing import Tuple
-
+ 
  from . import interleave_ffma
 diff --git a/deep_gemm/jit/interleave_ffma.py b/deep_gemm/jit/interleave_ffma.py
 index fcb377e..db9d6f3 100644
@@ -65,8 +65,8 @@ index fcb377e..db9d6f3 100644
  import subprocess
 -from torch.utils.cpp_extension import CUDA_HOME
 +from ..paddle_utils import CUDA_HOME
-
-
+ 
+ 
  def run_cuobjdump(file_path):
 diff --git a/deep_gemm/jit/runtime.py b/deep_gemm/jit/runtime.py
 index 66c370a..4761426 100644
@@ -78,7 +78,7 @@ index 66c370a..4761426 100644
 -import torch
 +import paddle
  from typing import Optional
-
+ 
  from .template import map_ctype
 @@ -35,7 +35,7 @@ class Runtime:
          assert len(args) == len(self.args), f'Expected {len(self.args)} arguments, got {len(args)}'
@@ -100,8 +100,8 @@ index ead37f5..51b02c1 100644
 -import torch
 +import paddle
  from typing import Any, Dict, Iterable, Tuple
-
-
+ 
+ 
  # Name map for Python `eval`
  typename_map: Dict[Any, str] = {
      **{t: t.__name__ for t in (bool, int, float)},
@@ -116,15 +116,15 @@ index ead37f5..51b02c1 100644
 +    paddle.float8_e4m3fn: 'paddle.float8_e4m3fn',
 +    paddle.device.cuda.Stream: "paddle.device.cuda.Stream",
  }
-
+ 
  # `ctype` map for Python casting
  ctype_map: Dict[Any, Any] = {
      **{t: getattr(ctypes, f'c_{t.__name__}') for t in (bool, int, float)},
 -    **{t: ctypes.c_void_p for t in (torch.int, torch.float, torch.bfloat16, torch.float8_e4m3fn, torch.cuda.Stream)},
 +    **{t: ctypes.c_void_p for t in (paddle.int32, paddle.float32, paddle.bfloat16, paddle.float8_e4m3fn, paddle.device.cuda.Stream)},
  }
-
-
+ 
+ 
 @@ -27,25 +27,25 @@ genc_map = {
      bool: ('bool', 'bool'),
      int: ('int', 'int'),
@@ -140,8 +140,8 @@ index ead37f5..51b02c1 100644
 +    paddle.float8_e4m3fn: ('void*', '__nv_fp8_e4m3*'),
 +    paddle.device.cuda.Stream: ('void*', 'cudaStream_t'),
  }
-
-
+ 
+ 
  def map_ctype(value: Any) -> Any:
      if hasattr(value, 'data_ptr'):
 -        if value.dtype == torch.int:
@@ -171,11 +171,11 @@ index cb438b7..44aa0ed 100644
 +import paddle
  from functools import lru_cache
  from typing import Tuple
-
+ 
 @@ -166,20 +166,20 @@ def get_best_configs(m: int, n: int, k: int, num_groups: int, num_sms: int,
      return num_min_sms, best_block_m, best_block_n, best_num_stages, best_tma_multicast_config, best_smem_config
-
-
+ 
+ 
 -def gemm_fp8_fp8_bf16_nt(lhs: Tuple[torch.Tensor, torch.Tensor],
 -                         rhs: Tuple[torch.Tensor, torch.Tensor],
 -                         out: torch.Tensor) -> None:
@@ -189,7 +189,7 @@ index cb438b7..44aa0ed 100644
      The LHS scaling tensor requires TMA-aligned transposed format, if your input does not match the requirement,
 -        this function will do a transposing with a set of slow PyTorch operations.
 +        this function will do a transposing with a set of slow paddle operations.
-
+ 
      Arguments:
 -        lhs: the first element is an FP8 tensor (typed `torch.float8_e4m3fn`) of shape `[m, k]`,
 +        lhs: the first element is an FP8 tensor (typed `paddle.float8_e4m3fn`) of shape `[m, k]`,
@@ -202,10 +202,10 @@ index cb438b7..44aa0ed 100644
 @@ -189,22 +189,22 @@ def gemm_fp8_fp8_bf16_nt(lhs: Tuple[torch.Tensor, torch.Tensor],
      n, k_ = rhs.shape
      m_, n_ = out.shape
-
+ 
 -    assert n % 64 == 0 and k % 128 == 0
 +    # assert n % 64 == 0 and k % 128 == 0
-
+ 
      # Type and shape checks
 -    assert m == m_ and n == n_ and k == k_
 -    assert n > 0 and k > 0
@@ -223,13 +223,13 @@ index cb438b7..44aa0ed 100644
 +    # assert rhs.dtype == paddle.float8_e4m3fn and rhs_scales.dtype == paddle.float32
 +    # assert out.dtype == paddle.bfloat16
 +    # assert lhs.is_contiguous() and rhs.is_contiguous() and out.is_contiguous()
-
+ 
      # LHS scales must be transposed for TMA load, but not for RHS scales
      # NOTES: `get_tma_aligned_lhs_scales` may launch a kernel if not processed by previous kernels
      lhs_scales = get_col_major_tma_aligned_tensor(lhs_scales)
 -    assert rhs_scales.is_contiguous()
 +    # assert rhs_scales.is_contiguous()
-
+ 
      # Do nothing if `m` is zero
      if m == 0:
 @@ -214,7 +214,7 @@ def gemm_fp8_fp8_bf16_nt(lhs: Tuple[torch.Tensor, torch.Tensor],
@@ -264,12 +264,12 @@ index 3b518c9..ba776bd 100644
 -import torch
 +import paddle
  from typing import Tuple
-
+ 
  from .gemm import get_best_configs, get_block_n_padding_for_smem_d
 @@ -37,25 +37,25 @@ gemm_t::run(out, rhs_scales, grouped_layout,
  """
-
-
+ 
+ 
 -def m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(lhs: Tuple[torch.Tensor, torch.Tensor],
 -                                              rhs: Tuple[torch.Tensor, torch.Tensor],
 -                                              out: torch.Tensor, m_indices: torch.Tensor) -> None:
@@ -285,7 +285,7 @@ index 3b518c9..ba776bd 100644
 +        this function will do a transposing with a set of slow Pypaddle operations.
      On the M axis, inputs are grouped into several batches, of which batch sizes aligned to
          `get_m_alignment_for_contiguous_layout()` (128).
-
+ 
      Arguments:
 -        lhs: the first element is an FP8 tensor (typed `torch.float8_e4m3fn`) of shape `[m_sum, k]`,
 +        lhs: the first element is an FP8 tensor (typed `paddle.float8_e4m3fn`) of shape `[m_sum, k]`,
@@ -301,7 +301,7 @@ index 3b518c9..ba776bd 100644
              Values of `m_indices` in every-m-alignment-block must also be the same.
 @@ -68,19 +68,19 @@ def m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(lhs: Tuple[torch.Tensor, torch.Ten
      m__ = m_indices.numel()
-
+ 
      # Type and shape checks
 -    assert m == m_ == m__ and k == k_ and n == n_
 -    assert lhs_scales.shape == (m, (k + 127) // 128)
@@ -321,12 +321,12 @@ index 3b518c9..ba776bd 100644
 +    # assert m_indices.dtype == paddle.int32
 +    # assert lhs.is_contiguous() and rhs.is_contiguous()
 +    # assert out.is_contiguous() and m_indices.is_contiguous()
-
+ 
      # LHS scales must be transposed for TMA load, but not for RHS scales
      lhs_scales = get_col_major_tma_aligned_tensor(lhs_scales)
 -    assert rhs_scales.is_contiguous()
 +    # assert rhs_scales.is_contiguous()
-
+ 
      # Do nothing if `m` is zero
      if m == 0:
 @@ -92,7 +92,7 @@ def m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(lhs: Tuple[torch.Tensor, torch.Ten
@@ -357,8 +357,8 @@ index 3b518c9..ba776bd 100644
      )
 @@ -118,22 +118,22 @@ def m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(lhs: Tuple[torch.Tensor, torch.Ten
      runtime(*args)
-
-
+ 
+ 
 -def m_grouped_gemm_fp8_fp8_bf16_nt_masked(lhs: Tuple[torch.Tensor, torch.Tensor],
 -                                          rhs: Tuple[torch.Tensor, torch.Tensor],
 -                                          out: torch.Tensor, masked_m: torch.Tensor, expected_m: int) -> None:
@@ -374,7 +374,7 @@ index 3b518c9..ba776bd 100644
 +        this function will do a transposing with a set of slow paddle operations.
      Moreover, this alignment requirement is different with the contiguous-format kernel, as we require that each batch
          should be separately transposed.
-
+ 
      Arguments:
 -        lhs: the first element is an FP8 tensor (typed `torch.float8_e4m3fn`) of shape `[num_groups, m_max, k]`,
 +        lhs: the first element is an FP8 tensor (typed `paddle.float8_e4m3fn`) of shape `[num_groups, m_max, k]`,
@@ -386,7 +386,7 @@ index 3b518c9..ba776bd 100644
          masked_m: a tensor of shape `[num_groups]`, `masked_m[i]` records actual rows of the `lhs[i]` matrix to compute
 @@ -149,21 +149,21 @@ def m_grouped_gemm_fp8_fp8_bf16_nt_masked(lhs: Tuple[torch.Tensor, torch.Tensor]
      num_groups___ = masked_m.numel()
-
+ 
      # Type and shape checks
 -    assert num_groups == num_groups_ == num_groups__ == num_groups___
 -    assert m == m_ and n == n_ and k == k_
@@ -410,16 +410,16 @@ index 3b518c9..ba776bd 100644
 +    # assert masked_m.dtype == paddle.int32
 +    # assert lhs.is_contiguous() and rhs.is_contiguous()
 +    # assert out.is_contiguous() and masked_m.is_contiguous()
-
+ 
      # LHS scales must be transposed for TMA load, but not for RHS scales
      lhs_scales = get_col_major_tma_aligned_tensor(lhs_scales)
 -    assert rhs_scales.is_contiguous()
 +    # assert rhs_scales.is_contiguous()
-
+ 
      # Auto-tuning with compilation
      global includes, template
 @@ -176,7 +176,7 @@ def m_grouped_gemm_fp8_fp8_bf16_nt_masked(lhs: Tuple[torch.Tensor, torch.Tensor]
-
+ 
      args = (lhs, lhs_scales, rhs, rhs_scales, out,
              masked_m, m,
 -            torch.cuda.current_stream(), num_sms, smem_config[0])
@@ -454,11 +454,11 @@ index 6ed6749..9e1d70f 100644
 -import torch
 +import paddle
  from typing import Any, Dict
-
+ 
  from ..jit import build, cpp_format, generate, Runtime
 @@ -51,10 +51,10 @@ class JITTuner:
                      continue
-
+ 
                  # Measure performance with L2 flush and a large GEMM kernel before to reduce overhead between kernels
 -                start_event = torch.cuda.Event(enable_timing=True)
 -                end_event = torch.cuda.Event(enable_timing=True)
@@ -478,9 +478,9 @@ index c6da56b..a17b1b1 100644
 @@ -1,4 +1,4 @@
 -import torch
 +import paddle
-
+ 
  _num_sms = None
-
+ 
 @@ -11,7 +11,7 @@ def set_num_sms(num_sms: int) -> None:
          num_sms: the desired maximum SM count for all GEMM kernels to use.
      """
@@ -488,8 +488,8 @@ index c6da56b..a17b1b1 100644
 -    assert 0 < num_sms <= torch.cuda.get_device_properties(device='cuda').multi_processor_count
 +    assert 0 < num_sms <= paddle.device.cuda.get_device_properties().multi_processor_count
      _num_sms = num_sms
-
-
+ 
+ 
 @@ -25,7 +25,7 @@ def get_num_sms() -> int:
      """
      global _num_sms
@@ -497,12 +497,12 @@ index c6da56b..a17b1b1 100644
 -        _num_sms = torch.cuda.get_device_properties(device='cuda').multi_processor_count
 +        _num_sms = paddle.device.cuda.get_device_properties().multi_processor_count
      return _num_sms
-
-
+ 
+ 
 @@ -74,9 +74,9 @@ def get_tma_aligned_size(x: int, element_size: int) -> int:
      return ceil_div(x, alignment) * alignment
-
-
+ 
+ 
 -def get_col_major_tma_aligned_tensor(x: torch.Tensor) -> torch.Tensor:
 +def get_col_major_tma_aligned_tensor(x: paddle.Tensor) -> paddle.Tensor:
      """
@@ -510,7 +510,7 @@ index c6da56b..a17b1b1 100644
 +    Returns TMA-aligned transposed format of the input tensor. `paddle.transpose` will be called if necessary.
      If the input tensor is already column-major layout and 16-byte aligned along the M axis
          (thus meets the requirement of LHS scaling tensor in DeepGEMM), this function will do nothing.
-
+ 
 @@ -92,18 +92,20 @@ def get_col_major_tma_aligned_tensor(x: torch.Tensor) -> torch.Tensor:
      m, n = x.shape[-2], x.shape[-1]
      aligned_m = get_tma_aligned_size(m, x.element_size())
@@ -519,14 +519,14 @@ index c6da56b..a17b1b1 100644
 +        if x.strides[0] == 1 and x.strides[1] == aligned_m:
              return x
          x, remove_dim = x.unsqueeze(0), True
-
+ 
      b = x.shape[0]
-
+ 
      # The last kernel gives a column-major TMA aligned layout
 -    if x.stride(0) == aligned_m * n and x.stride(1) == 1 and x.stride(2) == aligned_m:
 +    if x.strides[0] == aligned_m * n and x.strides[1] == 1 and x.strides[2] == aligned_m:
          return x.squeeze(0) if remove_dim else x
-
+ 
      # Normal layout requires transposing
 -    aligned_x = torch.transpose(torch.empty((b, n, aligned_m), device=x.device, dtype=x.dtype), 1, 2)
 +    aligned_x = paddle.transpose(
@@ -574,20 +574,20 @@ index d5cdd01..5237f09 100644
 -import torch.distributed as dist
 +import paddle
 +import paddle.distributed as dist
-
-
+ 
+ 
  def bench(fn, num_warmups: int = 5, num_tests: int = 10,
            high_precision: bool = False):
      # Flush L2 cache with 256 MB data
 -    torch.cuda.synchronize()
 -    cache = torch.empty(int(256e6 // 4), dtype=torch.int, device='cuda')
-+    paddle.device.synchronize()
++    paddle.device.cuda.synchronize()
 +    cache = paddle.empty((int(256e6 // 4)), dtype=paddle.int32)
      cache.zero_()
-
+ 
      # Warmup
 @@ -18,18 +18,18 @@ def bench(fn, num_warmups: int = 5, num_tests: int = 10,
-
+ 
      # Add a large kernel to eliminate the CPU launch overhead
      if high_precision:
 -        x = torch.randn((8192, 8192), dtype=torch.float, device='cuda')
@@ -595,7 +595,7 @@ index d5cdd01..5237f09 100644
 +        x = paddle.randn((8192, 8192), dtype=paddle.float32)
 +        y = paddle.randn((8192, 8192), dtype=paddle.float32)
          x @ y
-
+ 
      # Testing
 -    start_event = torch.cuda.Event(enable_timing=True)
 -    end_event = torch.cuda.Event(enable_timing=True)
@@ -607,9 +607,9 @@ index d5cdd01..5237f09 100644
      end_event.record()
 -    torch.cuda.synchronize()
 +    paddle.device.synchronize()
-
+ 
      return start_event.elapsed_time(end_event) / num_tests
-
+ 
 @@ -106,21 +106,21 @@ def bench_kineto(fn, kernel_names, num_tests: int = 30, suppress_kineto_output:
      # Profile
      suppress = suppress_stdout_stderr if suppress_kineto_output and not using_nsys else empty_suppress
@@ -636,7 +636,8 @@ index d5cdd01..5237f09 100644
 -                        torch.empty(flush_l2_size, dtype=torch.int, device='cuda').zero_()
 +                        paddle.empty(flush_l2_size, dtype=paddle.int32).zero_()
                      fn()
-
+ 
                  if not using_nsys:
---
+-- 
 2.43.0
+

custom_ops/cpu_ops/get_padding_offset.cc

@@ -84,6 +84,7 @@ std::vector<paddle::Tensor> GetPaddingOffset(const paddle::Tensor &input_ids,
                    seq_length,
                    bsz);
     return {x_remove_padding,
+            cum_offsets_out,
             padding_offset,
             cu_seqlens_q,
             cu_seqlens_k};
@@ -96,7 +97,7 @@ std::vector<std::vector<int64_t>> GetPaddingOffsetInferShape(
     const std::vector<int64_t> &seq_len_shape) {
     int64_t bsz = seq_len_shape[0];
     int64_t seq_len = input_ids_shape[1];
-    return {{-1}, {-1}, {bsz + 1}, {bsz + 1}};
+    return {{-1}, {bsz}, {-1}, {bsz + 1}, {bsz + 1}};
 }
 
 std::vector<paddle::DataType> GetPaddingOffsetInferDtype(
@@ -105,6 +106,7 @@ std::vector<paddle::DataType> GetPaddingOffsetInferDtype(
     const paddle::DataType &token_num_dtype,
     const paddle::DataType &seq_len_dtype) {
     return {input_ids_dtype,
+            seq_len_dtype,
             seq_len_dtype,
             seq_len_dtype,
             seq_len_dtype};
@@ -113,6 +115,7 @@ std::vector<paddle::DataType> GetPaddingOffsetInferDtype(
 PD_BUILD_STATIC_OP(get_padding_offset_cpu)
     .Inputs({"input_ids", "cum_offsets", "token_num", "seq_len"})
     .Outputs({"x_remove_padding",
+              "cum_offsets_out",
               "padding_offset",
               "cu_seqlens_q",
               "cu_seqlens_k"})

custom_ops/cpu_ops/rebuild_padding.cc

@@ -1,4 +1,4 @@
-// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
 // you may not use this file except in compliance with the License.
@@ -19,11 +19,10 @@
 #define PD_BUILD_STATIC_OP(name) PD_BUILD_OP(static_op_##name)
 #endif
 
-
 template <typename T>
 void RebuildPaddingCPUImpl(T *output_data,
                            const T *input_data,
-                           const int *cu_seqlens_q_data,
+                           const int *cum_offsets_data,
                            const int *seq_len_this_time_data,
                            const int *seq_lens_decoder_data,
                            const int *seq_lens_encoder_data,
@@ -41,12 +40,11 @@ void RebuildPaddingCPUImpl(T *output_data,
         if (seq_lens_decoder_data[bi] == 0 && seq_lens_encoder_data[bi] == 0) {
             continue;
         }
-
         if (seq_lens_encoder_data[bi] > 0) {
             seq_id = seq_lens_encoder_data[bi] - 1;
         }
-
-        const int ori_token_idx = cu_seqlens_q_data[bi] + seq_id;
+        const int ori_token_idx =
+            bi * max_input_length - cum_offsets_data[bi] + seq_id;
         const int src_offset = ori_token_idx * dim_embed + bias_idx;
 
         output_data[i] = input_data[src_offset];
@@ -56,7 +54,7 @@ void RebuildPaddingCPUImpl(T *output_data,
 template <typename T>
 void RebuildAppendPaddingCPUImpl(T *output_data,
                                  const T *input_data,
-                                 const int *cu_seqlens_q_data,
+                                 const int *cum_offsets_data,
                                  const int *seq_len_this_time_data,
                                  const int *seq_lens_decoder_data,
                                  const int *seq_lens_encoder_data,
@@ -71,32 +69,30 @@ void RebuildAppendPaddingCPUImpl(T *output_data,
         int bi = ori_token_id / max_input_length;
         if (seq_len_this_time_data[bi] == 0 ||
             (seq_lens_decoder_data[bi] == 0 &&
-            seq_lens_encoder_data[bi] == 0)) {
-                continue;
-            }
+             seq_lens_encoder_data[bi] == 0)) {
+            continue;
+        }
         int seq_id = 0;
-
         if (seq_lens_encoder_data[bi] > 0) {
             seq_id = seq_lens_encoder_data[bi] - 1;
         }
-        int input_token_id = cu_seqlens_q_data[bi] + seq_id;
+        int input_token_id = ori_token_id - cum_offsets_data[bi] + seq_id;
         int bias_idx = i % dim_embed;
         int src_offset = input_token_id * dim_embed + bias_idx;
-
         output_data[i] = input_data[src_offset];
     }
 }
 
 std::vector<paddle::Tensor> RebuildPaddingCPU(
     const paddle::Tensor &tmp_out,
-    const paddle::Tensor &cu_seqlens_q,
+    const paddle::Tensor &cum_offsets,
     const paddle::Tensor &seq_len_this_time,
     const paddle::Tensor &seq_lens_decoder,
     const paddle::Tensor &seq_lens_encoder,
     const paddle::optional<paddle::Tensor> &output_padding_offset,
     int max_input_length) {
     auto tmp_out_cpu = tmp_out.copy_to(paddle::CPUPlace(), true);
-    auto cu_seqlens_q_cpu = cu_seqlens_q.copy_to(paddle::CPUPlace(), true);
+    auto cum_offsets_cpu = cum_offsets.copy_to(paddle::CPUPlace(), true);
     auto seq_len_this_time_cpu =
         seq_len_this_time.copy_to(paddle::CPUPlace(), true);
     auto seq_lens_decoder_cpu =
@@ -111,7 +107,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
 
     int token_num = tmp_out_cpu.shape()[0];
     int dim_embed = tmp_out_cpu.shape()[1];
-    int bsz = cu_seqlens_q_cpu.shape()[0] - 1;
+    int bsz = cum_offsets_cpu.shape()[0];
 
     paddle::Tensor out;
     if (output_padding_offset_cpu) {
@@ -132,7 +128,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
             {bsz, dim_embed}, 0, tmp_out_cpu.dtype(), paddle::CPUPlace());
     }
 
-    const int *cu_seqlens_q_data = cu_seqlens_q_cpu.data<int>();
+    const int *cum_offsets_data = cum_offsets_cpu.data<int>();
     const int *seq_len_this_time_data = seq_len_this_time_cpu.data<int>();
     const int *seq_lens_decoder_data = seq_lens_decoder_cpu.data<int>();
     const int *seq_lens_encoder_data = seq_lens_encoder_cpu.data<int>();
@@ -145,7 +141,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
             case paddle::DataType::FLOAT32:
                 RebuildAppendPaddingCPUImpl<float>(out.data<float>(),
                                                    tmp_out_cpu.data<float>(),
-                                                   cu_seqlens_q_data,
+                                                   cum_offsets_data,
                                                    seq_len_this_time_data,
                                                    seq_lens_decoder_data,
                                                    seq_lens_encoder_data,
@@ -158,7 +154,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
                 RebuildAppendPaddingCPUImpl<paddle::float16>(
                     out.data<paddle::float16>(),
                     tmp_out_cpu.data<paddle::float16>(),
-                    cu_seqlens_q_data,
+                    cum_offsets_data,
                     seq_len_this_time_data,
                     seq_lens_decoder_data,
                     seq_lens_encoder_data,
@@ -171,7 +167,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
                 RebuildAppendPaddingCPUImpl<paddle::bfloat16>(
                     out.data<paddle::bfloat16>(),
                     tmp_out_cpu.data<paddle::bfloat16>(),
-                    cu_seqlens_q_data,
+                    cum_offsets_data,
                     seq_len_this_time_data,
                     seq_lens_decoder_data,
                     seq_lens_encoder_data,
@@ -190,7 +186,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
             case paddle::DataType::FLOAT32:
                 RebuildPaddingCPUImpl<float>(out.data<float>(),
                                              tmp_out_cpu.data<float>(),
-                                             cu_seqlens_q_data,
+                                             cum_offsets_data,
                                              seq_len_this_time_data,
                                              seq_lens_decoder_data,
                                              seq_lens_encoder_data,
@@ -202,7 +198,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
                 RebuildPaddingCPUImpl<paddle::float16>(
                     out.data<paddle::float16>(),
                     tmp_out_cpu.data<paddle::float16>(),
-                    cu_seqlens_q_data,
+                    cum_offsets_data,
                     seq_len_this_time_data,
                     seq_lens_decoder_data,
                     seq_lens_encoder_data,
@@ -211,10 +207,11 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
                     elem_nums);
                 break;
             case paddle::DataType::BFLOAT16:
+
                 RebuildPaddingCPUImpl<paddle::bfloat16>(
                     out.data<paddle::bfloat16>(),
                     tmp_out_cpu.data<paddle::bfloat16>(),
-                    cu_seqlens_q_data,
+                    cum_offsets_data,
                     seq_len_this_time_data,
                     seq_lens_decoder_data,
                     seq_lens_encoder_data,
@@ -233,7 +230,7 @@ std::vector<paddle::Tensor> RebuildPaddingCPU(
 
 std::vector<std::vector<int64_t>> RebuildPaddingInferShape(
     const std::vector<int64_t> &tmp_out_shape,
-    const std::vector<int64_t> &cu_seqlens_q_shape,
+    const std::vector<int64_t> &cum_offsets_shape,
     const std::vector<int64_t> &seq_len_this_time_shape,
     const std::vector<int64_t> &seq_lens_decoder_shape,
     const std::vector<int64_t> &seq_lens_encoder_shape,
@@ -242,14 +239,14 @@ std::vector<std::vector<int64_t>> RebuildPaddingInferShape(
     if (output_padding_offset_shape) {
         return {{-1, dim_embed}};
     } else {
-        int64_t bsz = cu_seqlens_q_shape[0] - 1;
+        int64_t bsz = cum_offsets_shape[0];
         return {{bsz, dim_embed}};
     }
 }
 
 std::vector<paddle::DataType> RebuildPaddingInferDtype(
     const paddle::DataType &tmp_out_dtype,
-    const paddle::DataType &cu_seqlens_q_dtype,
+    const paddle::DataType &cum_offsets_dtype,
     const paddle::DataType &seq_len_this_time_dtype,
     const paddle::DataType &seq_lens_decoder_dtype,
     const paddle::DataType &seq_lens_encoder_dtype,
@@ -259,7 +256,7 @@ std::vector<paddle::DataType> RebuildPaddingInferDtype(
 
 PD_BUILD_STATIC_OP(rebuild_padding_cpu)
     .Inputs({"tmp_out",
-             "cu_seqlens_q",
+             "cum_offsets",
              "seq_len_this_time",
              "seq_lens_decoder",
              "seq_lens_encoder",

custom_ops/gpu_ops/append_attention.cu

@@ -38,7 +38,7 @@ class type2value<phi::dtype::float16> {
 
 
 template <paddle::DataType D>
-void AppendAttentionKernel(
+std::vector<paddle::Tensor> AppendAttentionKernel(
     const AppendAttnMetaData& meta_data,
     const paddle::Tensor& qkv,
     const paddle::Tensor& key_cache,
@@ -46,8 +46,8 @@ void AppendAttentionKernel(
     const paddle::Tensor& seq_lens_encoder,
     const paddle::Tensor& seq_lens_decoder,
     const paddle::Tensor& seq_lens_this_time,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::Tensor& encoder_batch_ids,
     const paddle::Tensor& encoder_tile_ids_per_batch,
@@ -60,7 +60,6 @@ void AppendAttentionKernel(
     const paddle::Tensor& decoder_num_blocks,
     const paddle::Tensor& set_max_lengths,
     const paddle::Tensor& max_len_kv,
-    paddle::Tensor& fmha_out,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& attn_mask,
     const paddle::optional<paddle::Tensor>& qkv_bias,
@@ -73,11 +72,7 @@ void AppendAttentionKernel(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const paddle::optional<paddle::Tensor>& out_linear_shifts,
     const paddle::optional<paddle::Tensor>& out_linear_smooths,
-    const paddle::optional<paddle::Tensor>& mask_offset,
     const paddle::optional<paddle::Tensor>& kv_signal_data,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
     const bool rope_3d,
@@ -123,6 +118,27 @@ void AppendAttentionKernel(
   } else {
     qkv_out = qkv;
   }
+  paddle::Tensor fmha_out;
+  if (out_linear_in_scale > 0.0) {
+    if (fabs(quant_max_bound - 127.0f) < 0.000001) {
+      fmha_out = GetEmptyTensor(
+        {meta_data.token_nums, meta_data.q_num_heads * meta_data.head_dims},
+        paddle::DataType::INT8,
+        qkv.place());
+    } else if (fabs(quant_max_bound - 448.0f) < 0.000001) {
+      fmha_out = GetEmptyTensor(
+        {meta_data.token_nums, meta_data.q_num_heads * meta_data.head_dims},
+        paddle::DataType::FLOAT8_E4M3FN,
+        qkv.place());
+    }else{
+      PD_THROW("Only supported attr of quant_max_bound in ['127', '448'].");
+    }
+  } else {
+    fmha_out = GetEmptyTensor(
+        {meta_data.token_nums, meta_data.q_num_heads * meta_data.head_dims},
+        D,
+        qkv.place());
+  }
 
   auto dispatch_CascadeAppendAttentionKernel = [&](auto temp_args,
         const paddle::Tensor& lambda_batch_ids,
@@ -140,8 +156,8 @@ void AppendAttentionKernel(
           key_cache,
           value_cache,
           attn_mask,
-          cache_quant_type_str == "block_wise_fp8" ? cache_k_quant_scales : cache_k_dequant_scales,
-          cache_quant_type_str == "block_wise_fp8" ? cache_v_quant_scales : cache_v_dequant_scales,
+          cache_k_dequant_scales,
+          cache_v_dequant_scales,
           cache_k_zp,
           cache_v_zp,
           out_linear_shifts,
@@ -149,8 +165,8 @@ void AppendAttentionKernel(
           seq_lens_this_time,
           seq_lens_decoder,
           seq_lens_encoder,
-          batch_id_per_token,
-          cu_seqlens_q,
+          padding_offsets,
+          cum_offsets,
           block_tables,
           lambda_batch_ids,
           lambda_tile_ids_per_batch,
@@ -186,8 +202,8 @@ void AppendAttentionKernel(
           seq_lens_this_time,
           seq_lens_encoder,
           seq_lens_decoder,
-          batch_id_per_token,
-          cu_seqlens_q,
+          padding_offsets,
+          cum_offsets,
           block_tables,
           kv_batch_ids,
           kv_tile_ids_per_batch,
@@ -207,10 +223,7 @@ void AppendAttentionKernel(
           main_stream,
           &qkv_out,
           const_cast<paddle::Tensor*>(&key_cache),
-          const_cast<paddle::Tensor*>(&value_cache),
-          q_norm_weight,
-          k_norm_weight,
-          rms_norm_eps);
+          const_cast<paddle::Tensor*>(&value_cache));
     };
 
     if (qkv_out_scales) {
@@ -261,8 +274,8 @@ void AppendAttentionKernel(
             qkv,  // [token_num, num_heads, head_dim]
             seq_lens_decoder,
             seq_lens_encoder,
-            batch_id_per_token,
-            cu_seqlens_q,
+            padding_offsets,
+            cum_offsets,
             block_tables,
             rotary_embs,
             qkv_out_scales,
@@ -273,23 +286,19 @@ void AppendAttentionKernel(
             cache_v_zp,
             cache_quant_type_str,
             use_neox_rotary_style,
-            rope_3d,
             max_input_length,
             exec_stream,
             &qkv_out,
             const_cast<paddle::Tensor*>(&key_cache),
-            const_cast<paddle::Tensor*>(&value_cache),
-            q_norm_weight,
-            k_norm_weight,
-            rms_norm_eps);
+            const_cast<paddle::Tensor*>(&value_cache));
       } else {
         SpeculateWriteCacheWithRoPEKernel<data_t, data_t>(
             meta_data,
             qkv_out,  // [token_num, num_heads, head_dim]
             seq_lens_decoder,
             seq_lens_encoder,
-            batch_id_per_token,
-            cu_seqlens_q,
+            padding_offsets,
+            cum_offsets,
             block_tables,
             rotary_embs,
             qkv_out_scales,
@@ -300,15 +309,11 @@ void AppendAttentionKernel(
             cache_v_zp,
             cache_quant_type_str,
             use_neox_rotary_style,
-            rope_3d,
             max_input_length,
             exec_stream,
             &qkv_out,
             const_cast<paddle::Tensor*>(&key_cache),
-            const_cast<paddle::Tensor*>(&value_cache),
-            q_norm_weight,
-            k_norm_weight,
-            rms_norm_eps);
+            const_cast<paddle::Tensor*>(&value_cache));
       }
     } else {
       if (qkv_out_scales) {
@@ -317,8 +322,8 @@ void AppendAttentionKernel(
             qkv,  // [token_num, num_heads, head_dim]
             seq_lens_decoder,
             seq_lens_encoder,
-            batch_id_per_token,
-            cu_seqlens_q,
+            padding_offsets,
+            cum_offsets,
             block_tables,
             rotary_embs,
             qkv_out_scales,
@@ -334,18 +339,15 @@ void AppendAttentionKernel(
             exec_stream,
             &qkv_out,
             const_cast<paddle::Tensor*>(&key_cache),
-            const_cast<paddle::Tensor*>(&value_cache),
-            q_norm_weight,
-            k_norm_weight,
-            rms_norm_eps);
+            const_cast<paddle::Tensor*>(&value_cache));
       } else {
         DecoderWriteCacheWithRoPEKernel<data_t, data_t>(
             meta_data,
             qkv_out,  // [token_num, num_heads, head_dim]
             seq_lens_decoder,
             seq_lens_encoder,
-            batch_id_per_token,
-            cu_seqlens_q,
+            padding_offsets,
+            cum_offsets,
             block_tables,
             rotary_embs,
             qkv_out_scales,
@@ -361,10 +363,7 @@ void AppendAttentionKernel(
             exec_stream,
             &qkv_out,
             const_cast<paddle::Tensor*>(&key_cache),
-            const_cast<paddle::Tensor*>(&value_cache),
-            q_norm_weight,
-            k_norm_weight,
-            rms_norm_eps);
+            const_cast<paddle::Tensor*>(&value_cache));
       }
     }
 
@@ -393,6 +392,8 @@ void AppendAttentionKernel(
       cudaStreamWaitEvent(main_stream, decoder_event);
     }
   }
+
+  return {fmha_out, qkv_out};
 }
 
 std::vector<paddle::Tensor> AppendAttention(
@@ -402,8 +403,8 @@ std::vector<paddle::Tensor> AppendAttention(
     const paddle::Tensor& seq_lens_encoder,
     const paddle::Tensor& seq_lens_decoder,
     const paddle::Tensor& seq_lens_this_time,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::Tensor& encoder_batch_ids,
     const paddle::Tensor& encoder_tile_ids_per_batch,
@@ -428,11 +429,7 @@ std::vector<paddle::Tensor> AppendAttention(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const paddle::optional<paddle::Tensor>& out_linear_shifts,
     const paddle::optional<paddle::Tensor>& out_linear_smooths,
-    const paddle::optional<paddle::Tensor>& mask_offset,
     const paddle::optional<paddle::Tensor>& kv_signal_data,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps,
     const std::string& compute_dtype,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
@@ -465,62 +462,10 @@ std::vector<paddle::Tensor> AppendAttention(
 
   meta_data.max_blocks_per_seq = block_tables.dims()[1];
   meta_data.block_size = key_cache.dims()[2];
-  meta_data.batch_size = seq_lens_this_time.dims()[0];
+  meta_data.batch_size = cum_offsets.dims()[0];
 
-  // template dtype generation
-  phi::DataType dtype_id;
-  switch (qkv.dtype()) {
-    case paddle::DataType::FLOAT16:  {dtype_id = phi::DataType::FLOAT16;  break;}
-    case paddle::DataType::BFLOAT16: {dtype_id = phi::DataType::BFLOAT16; break;}
-    case paddle::DataType::INT32: {
-      if (compute_dtype == "bf16") {
-        dtype_id = phi::DataType::BFLOAT16;
-        break;
-      } else if (compute_dtype == "fp16") {
-        dtype_id = phi::DataType::FLOAT16;
-        break;
-      } else {
-        PD_THROW("Only supported attr of compute_dtype in ['fp16', 'bf16'].");
-        break;
-      }
-    }
-    default: {
-      PD_THROW(
-          "NOT supported data type. "
-          "Only float16 and bfloat16 are supported. ");
-      break;
-    }
-  }
-
-  // fmha_out generation, rewrite from AppendAttentionKernel
-  paddle::Tensor fmha_out;
-  if (out_linear_in_scale > 0.0) {
-    if (fabs(quant_max_bound - 127.0f) < 0.000001) {
-      fmha_out = GetEmptyTensor(
-        {meta_data.token_nums, meta_data.q_num_heads * meta_data.head_dims},
-        paddle::DataType::INT8,
-        qkv.place());
-    } else if (fabs(quant_max_bound - 448.0f) < 0.000001) {
-      fmha_out = GetEmptyTensor(
-        {meta_data.token_nums, meta_data.q_num_heads * meta_data.head_dims},
-        paddle::DataType::FLOAT8_E4M3FN,
-        qkv.place());
-    } else{
-      PD_THROW("Only supported attr of quant_max_bound in ['127', '448'].");
-    }
-  } else {
-    fmha_out = GetEmptyTensor(
-        {meta_data.token_nums, meta_data.q_num_heads * meta_data.head_dims},
-        dtype_id,
-        qkv.place());
-  }
-
-  if (mask_offset) {
-    meta_data.mask_offset = mask_offset.get().data<int>();
-  }
-
-  auto dispatch_by_template = [&](auto temp_args) -> void {
-      AppendAttentionKernel<type2value<decltype(temp_args)>::value>(
+  auto dispatch_by_template = [&](auto temp_args) -> std::vector<paddle::Tensor> {
+      return AppendAttentionKernel<type2value<decltype(temp_args)>::value>(
           meta_data,
           qkv,
           key_cache,
@@ -528,8 +473,8 @@ std::vector<paddle::Tensor> AppendAttention(
           seq_lens_encoder,
           seq_lens_decoder,
           seq_lens_this_time,
-          batch_id_per_token,
-          cu_seqlens_q,
+          padding_offsets,
+          cum_offsets,
           block_tables,
           encoder_batch_ids,
           encoder_tile_ids_per_batch,
@@ -542,7 +487,6 @@ std::vector<paddle::Tensor> AppendAttention(
           decoder_num_blocks,
           set_max_lengths,
           max_len_kv,
-          fmha_out,
           rotary_embs,
           attn_mask,
           qkv_bias,
@@ -555,11 +499,7 @@ std::vector<paddle::Tensor> AppendAttention(
           cache_v_zp,
           out_linear_shifts,
           out_linear_smooths,
-          mask_offset,
           kv_signal_data,
-          q_norm_weight,
-          k_norm_weight,
-          rms_norm_eps,
           cache_quant_type_str,
           use_neox_rotary_style,
           rope_3d,
@@ -574,198 +514,35 @@ std::vector<paddle::Tensor> AppendAttention(
           speculate_max_draft_token_num,
           causal,
           speculate_decoder);
-  };
 
+  };
 
   phi::dtype::float16 fp16_dtype;
   phi::dtype::bfloat16 bp16_dtype;
-  switch (dtype_id){
-      case phi::DataType::FLOAT16: {
-        dispatch_by_template(fp16_dtype);
-        return {fmha_out};
-        }
-      case phi::DataType::BFLOAT16: {
-        dispatch_by_template(bp16_dtype);
-        return {fmha_out};
-        }
-      default:
-        PD_THROW(
+
+  switch (qkv.dtype()) {
+    case paddle::DataType::FLOAT16: return dispatch_by_template(fp16_dtype);
+    case paddle::DataType::BFLOAT16: return dispatch_by_template(bp16_dtype);
+    case paddle::DataType::INT32: {
+      if (compute_dtype == "bf16") {
+        return dispatch_by_template(bp16_dtype);
+      } else if (compute_dtype == "fp16") {
+        return dispatch_by_template(fp16_dtype);
+      } else {
+        PD_THROW("Only supported attr of compute_dtype in ['fp16', 'bf16'].");
+        break;
+      }
+    }
+    default: {
+      PD_THROW(
           "NOT supported data type. "
           "Only float16 and bfloat16 are supported. ");
-        break;
+      break;
+    }
   }
-
   return {paddle::Tensor{}};
 }
 
-void AppendAttentionWithOutput(
-    const paddle::Tensor& qkv,
-    const paddle::Tensor& key_cache,
-    const paddle::Tensor& value_cache,
-    const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& seq_lens_decoder,
-    const paddle::Tensor& seq_lens_this_time,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
-    const paddle::Tensor& block_tables,
-    const paddle::Tensor& encoder_batch_ids,
-    const paddle::Tensor& encoder_tile_ids_per_batch,
-    const paddle::Tensor& encoder_num_blocks,
-    const paddle::Tensor& kv_batch_ids,
-    const paddle::Tensor& kv_tile_ids_per_batch,
-    const paddle::Tensor& kv_num_blocks,
-    const paddle::Tensor& decoder_batch_ids,
-    const paddle::Tensor& decoder_tile_ids_per_batch,
-    const paddle::Tensor& decoder_num_blocks,
-    const paddle::Tensor& set_max_lengths,
-    const paddle::Tensor& max_len_kv,
-    paddle::Tensor& fmha_out,
-    const paddle::optional<paddle::Tensor>& rotary_embs,
-    const paddle::optional<paddle::Tensor>& attn_mask,
-    const paddle::optional<paddle::Tensor>& qkv_bias,
-    const paddle::optional<paddle::Tensor>& qkv_out_scales,
-    const paddle::optional<paddle::Tensor>& cache_k_quant_scales,
-    const paddle::optional<paddle::Tensor>& cache_v_quant_scales,
-    const paddle::optional<paddle::Tensor>& cache_k_dequant_scales,
-    const paddle::optional<paddle::Tensor>& cache_v_dequant_scales,
-    const paddle::optional<paddle::Tensor>& cache_k_zp,
-    const paddle::optional<paddle::Tensor>& cache_v_zp,
-    const paddle::optional<paddle::Tensor>& out_linear_shifts,
-    const paddle::optional<paddle::Tensor>& out_linear_smooths,
-    const paddle::optional<paddle::Tensor>& mask_offset,
-    const paddle::optional<paddle::Tensor>& kv_signal_data,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps,
-    const std::string& compute_dtype,
-    const std::string& cache_quant_type_str,
-    const bool use_neox_rotary_style,
-    const bool rope_3d,
-    const int max_input_length,
-    const float quant_max_bound,
-    const float quant_min_bound,
-    const float out_linear_in_scale,
-    const int encoder_block_shape_q,
-    const int decoder_block_shape_q,
-    const int max_partition_size,
-    const int encoder_max_partition_size,
-    const int speculate_max_draft_token_num,
-    const bool causal,
-    const bool speculate_decoder) {
-  AppendAttnMetaData meta_data;
-
-  const auto& qkv_dims = qkv.dims();
-  const auto& key_cache_dims = key_cache.dims();
-  meta_data.token_nums = qkv_dims[0];
-  meta_data.kv_num_heads = key_cache_dims[1];
-  meta_data.head_dims = key_cache_dims[3];
-  // TODO: trick method support c4, add attr head_dims in the future
-  if (cache_quant_type_str == "cache_int4_zp") {
-    meta_data.head_dims *= 2;
-  }
-  const int total_num_head =
-      qkv_dims[qkv_dims.size() - 1] / meta_data.head_dims;
-  meta_data.q_num_heads = total_num_head - 2 * meta_data.kv_num_heads;
-
-  meta_data.max_blocks_per_seq = block_tables.dims()[1];
-  meta_data.block_size = key_cache.dims()[2];
-  meta_data.batch_size = seq_lens_this_time.dims()[0];
-
-  if (mask_offset) {
-    meta_data.mask_offset = mask_offset.get().data<int>();
-  }
-
-  auto dispatch_by_template = [&](auto temp_args) -> void {
-      AppendAttentionKernel<type2value<decltype(temp_args)>::value>(
-          meta_data,
-          qkv,
-          key_cache,
-          value_cache,
-          seq_lens_encoder,
-          seq_lens_decoder,
-          seq_lens_this_time,
-          batch_id_per_token,
-          cu_seqlens_q,
-          block_tables,
-          encoder_batch_ids,
-          encoder_tile_ids_per_batch,
-          encoder_num_blocks,
-          kv_batch_ids,
-          kv_tile_ids_per_batch,
-          kv_num_blocks,
-          decoder_batch_ids,
-          decoder_tile_ids_per_batch,
-          decoder_num_blocks,
-          set_max_lengths,
-          max_len_kv,
-          fmha_out,
-          rotary_embs,
-          attn_mask,
-          qkv_bias,
-          qkv_out_scales,
-          cache_k_quant_scales,
-          cache_v_quant_scales,
-          cache_k_dequant_scales,
-          cache_v_dequant_scales,
-          cache_k_zp,
-          cache_v_zp,
-          out_linear_shifts,
-          out_linear_smooths,
-          mask_offset,
-          kv_signal_data,
-          q_norm_weight,
-          k_norm_weight,
-          rms_norm_eps,
-          cache_quant_type_str,
-          use_neox_rotary_style,
-          rope_3d,
-          max_input_length,
-          quant_max_bound,
-          quant_min_bound,
-          out_linear_in_scale,
-          encoder_block_shape_q,
-          decoder_block_shape_q,
-          max_partition_size,
-          encoder_max_partition_size,
-          speculate_max_draft_token_num,
-          causal,
-          speculate_decoder);
-  };
-
-  phi::dtype::float16 fp16_dtype;
-  phi::dtype::bfloat16 bp16_dtype;
-
-  switch (qkv.dtype()) {
-    case paddle::DataType::FLOAT16: {
-      dispatch_by_template(fp16_dtype);
-      break;
-    }
-    case paddle::DataType::BFLOAT16: {
-      dispatch_by_template(bp16_dtype);
-      break;
-    }
-    case paddle::DataType::INT32: {
-      if (compute_dtype == "bf16") {
-        dispatch_by_template(bp16_dtype);
-        break;
-      } else if (compute_dtype == "fp16") {
-        dispatch_by_template(fp16_dtype);
-        break;
-      } else {
-        PD_THROW("Only supported attr of compute_dtype in ['fp16', 'bf16'].");
-        break;
-      }
-    }
-    default: {
-      PD_THROW(
-          "NOT supported data type. "
-          "Only float16 and bfloat16 are supported. ");
-      break;
-    }
-  }
-}
-
-
 std::vector<std::vector<int64_t>> AppendAttentionInferShape(
     const std::vector<int64_t>& qkv_shape,
     const std::vector<int64_t>& key_cache_shape,
@@ -773,8 +550,8 @@ std::vector<std::vector<int64_t>> AppendAttentionInferShape(
     const std::vector<int64_t>& seq_lens_encoder_shape,
     const std::vector<int64_t>& seq_lens_decoder_shape,
     const std::vector<int64_t>& seq_lens_this_time_shape,
-    const std::vector<int64_t>& batch_id_per_token_shape,
-    const std::vector<int64_t>& cu_seqlens_q_shape,
+    const std::vector<int64_t>& padding_offsets_shape,
+    const std::vector<int64_t>& cum_offsets_shape,
     const std::vector<int64_t>& block_tables_shape,
     const std::vector<int64_t>& encoder_batch_ids_shape,
     const std::vector<int64_t>& encoder_tile_ids_per_batch_shape,
@@ -799,11 +576,7 @@ std::vector<std::vector<int64_t>> AppendAttentionInferShape(
     const paddle::optional<std::vector<int64_t>>& cache_v_zp_shape,
     const paddle::optional<std::vector<int64_t>>& out_linear_shifts_shape,
     const paddle::optional<std::vector<int64_t>>& out_linear_smooths_shape,
-    const paddle::optional<std::vector<int64_t>>& mask_offset_shape,
     const paddle::optional<std::vector<int64_t>>& kv_signal_data_shape,
-    const paddle::optional<std::vector<int64_t>>& q_norm_weight_shape,
-    const paddle::optional<std::vector<int64_t>>& k_norm_weight_shape,
-    const float rms_norm_eps,
     const std::string& compute_dtype,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
@@ -827,7 +600,7 @@ std::vector<std::vector<int64_t>> AppendAttentionInferShape(
   }
   const int total_num_head = qkv_shape[qkv_shape.size() - 1] / head_dim;
   const int num_heads = total_num_head - 2 * kv_num_heads;
-  return {{token_num, num_heads * head_dim}};
+  return {{token_num, num_heads * head_dim}, qkv_shape};
 }
 
 std::vector<paddle::DataType> AppendAttentionInferDtype(
@@ -837,8 +610,8 @@ std::vector<paddle::DataType> AppendAttentionInferDtype(
     const paddle::DataType& seq_lens_encoder_dtype,
     const paddle::DataType& seq_lens_decoder_dtype,
     const paddle::DataType& seq_lens_this_time_dtype,
-    const paddle::DataType& batch_id_per_token_dtype,
-    const paddle::DataType& cu_seqlens_q_dtype,
+    const paddle::DataType& padding_offsets_dtype,
+    const paddle::DataType& cum_offsets_dtype,
     const paddle::DataType& block_tables_dtype,
     const paddle::DataType& encoder_batch_ids_dtype,
     const paddle::DataType& encoder_tile_ids_per_batch_dtype,
@@ -863,11 +636,7 @@ std::vector<paddle::DataType> AppendAttentionInferDtype(
     const paddle::optional<paddle::DataType>& cache_v_zp_dtype,
     const paddle::optional<paddle::DataType>& out_linear_shifts_dtype,
     const paddle::optional<paddle::DataType>& out_linear_smooths_dtype,
-    const paddle::optional<paddle::DataType>& mask_offset_dtype,
     const paddle::optional<paddle::DataType>& kv_signal_data_dtype,
-    const paddle::optional<paddle::DataType>& q_norm_weight_dtype,
-    const paddle::optional<paddle::DataType>& k_norm_weight_dtype,
-    const float rms_norm_eps,
     const std::string& compute_dtype,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
@@ -886,148 +655,32 @@ std::vector<paddle::DataType> AppendAttentionInferDtype(
   if (compute_dtype == "bf16") {
     if (out_linear_in_scale > 0.0) {
       if (fabs(quant_max_bound - 127.0f) < 0.000001) {
-        return {paddle::DataType::INT8};
+        return {paddle::DataType::INT8, paddle::DataType::BFLOAT16};
       } else if (fabs(quant_max_bound - 448.0f) < 0.000001) {
-        return {paddle::DataType::FLOAT8_E4M3FN};
+        return {paddle::DataType::FLOAT8_E4M3FN, paddle::DataType::BFLOAT16};
       }else{
         PD_THROW("Only supported attr of quant_max_bound in ['127.0', '448.0'].");
       }
     } else {
-      return {paddle::DataType::BFLOAT16};
+      return {paddle::DataType::BFLOAT16, paddle::DataType::BFLOAT16};
     }
   } else if (compute_dtype == "fp16") {
     if (out_linear_in_scale > 0.0) {
       if (fabs(quant_max_bound - 127.0f) < 0.000001) {
-        return {paddle::DataType::INT8};
+        return {paddle::DataType::INT8, paddle::DataType::FLOAT16};
       } else if (fabs(quant_max_bound - 448.0f) < 0.000001) {
-        return {paddle::DataType::FLOAT8_E4M3FN};
+        return {paddle::DataType::FLOAT8_E4M3FN, paddle::DataType::FLOAT16};
       }else{
         PD_THROW("Only supported attr of quant_max_bound in ['127.0', '448.0'].");
       }
     } else {
-      return {paddle::DataType::FLOAT16};
+      return {paddle::DataType::FLOAT16, paddle::DataType::FLOAT16};
     }
   } else {
     PD_THROW("Only supported attr of compute_dtype in ['fp16', 'bf16'].");
   }
 }
 
-std::vector<std::vector<int64_t>> AppendAttentionWithOutputInferShape(
-    const std::vector<int64_t>& qkv_shape,
-    const std::vector<int64_t>& key_cache_shape,
-    const std::vector<int64_t>& value_cache_shape,
-    const std::vector<int64_t>& seq_lens_encoder_shape,
-    const std::vector<int64_t>& seq_lens_decoder_shape,
-    const std::vector<int64_t>& seq_lens_this_time_shape,
-    const std::vector<int64_t>& batch_id_per_token_shape,
-    const std::vector<int64_t>& cu_seqlens_q_shape,
-    const std::vector<int64_t>& block_tables_shape,
-    const std::vector<int64_t>& encoder_batch_ids_shape,
-    const std::vector<int64_t>& encoder_tile_ids_per_batch_shape,
-    const std::vector<int64_t>& encoder_num_blocks_shape,
-    const std::vector<int64_t>& kv_batch_ids_shape,
-    const std::vector<int64_t>& kv_tile_ids_per_batch_shape,
-    const std::vector<int64_t>& kv_num_blocks_shape,
-    const std::vector<int64_t>& decoder_batch_ids_shape,
-    const std::vector<int64_t>& decoder_tile_ids_per_batch_shape,
-    const std::vector<int64_t>& decoder_num_blocks_shape,
-    const std::vector<int64_t>& set_max_lengths_shape,
-    const std::vector<int64_t>& max_len_kv_shape,
-    const std::vector<int64_t>& fmha_out_shape,
-    const paddle::optional<std::vector<int64_t>>& rotary_embs_shape,
-    const paddle::optional<std::vector<int64_t>>& attn_mask_shape,
-    const paddle::optional<std::vector<int64_t>>& qkv_bias_shape,
-    const paddle::optional<std::vector<int64_t>>& qkv_out_scales_shape,
-    const paddle::optional<std::vector<int64_t>>& cache_k_quant_scales_shape,
-    const paddle::optional<std::vector<int64_t>>& cache_v_quant_scales_shape,
-    const paddle::optional<std::vector<int64_t>>& cache_k_dequant_scales_shape,
-    const paddle::optional<std::vector<int64_t>>& cache_v_dequant_scales_shape,
-    const paddle::optional<std::vector<int64_t>>& cache_k_zp_shape,
-    const paddle::optional<std::vector<int64_t>>& cache_v_zp_shape,
-    const paddle::optional<std::vector<int64_t>>& out_linear_shifts_shape,
-    const paddle::optional<std::vector<int64_t>>& out_linear_smooths_shape,
-    const paddle::optional<std::vector<int64_t>>& mask_offset_shape,
-    const paddle::optional<std::vector<int64_t>>& kv_signal_data_shape,
-    const paddle::optional<std::vector<int64_t>>& q_norm_weight_shape,
-    const paddle::optional<std::vector<int64_t>>& k_norm_weight_shape,
-    const float rms_norm_eps,
-    const std::string& compute_dtype,
-    const std::string& cache_quant_type_str,
-    const bool use_neox_rotary_style,
-    const bool rope_3d,
-    const int max_input_length,
-    const float quant_max_bound,
-    const float quant_min_bound,
-    const float out_linear_in_scale,
-    const int encoder_block_shape_q,
-    const int decoder_block_shape_q,
-    const int max_partition_size,
-    const int encoder_max_partition_size,
-    const int speculate_max_draft_token_num,
-    const bool causal,
-    const bool speculate_decoder) {
-  return {fmha_out_shape};
-}
-
-std::vector<paddle::DataType> AppendAttentionWithOutputInferDtype(
-    const paddle::DataType& qkv_dtype,
-    const paddle::DataType& key_cache_dtype,
-    const paddle::DataType& value_cache_dtype,
-    const paddle::DataType& seq_lens_encoder_dtype,
-    const paddle::DataType& seq_lens_decoder_dtype,
-    const paddle::DataType& seq_lens_this_time_dtype,
-    const paddle::DataType& batch_id_per_token_dtype,
-    const paddle::DataType& cu_seqlens_q_dtype,
-    const paddle::DataType& block_tables_dtype,
-    const paddle::DataType& encoder_batch_ids_dtype,
-    const paddle::DataType& encoder_tile_ids_per_batch_dtype,
-    const paddle::DataType& encoder_num_blocks_dtype,
-    const paddle::DataType& kv_batch_ids_dtype,
-    const paddle::DataType& kv_tile_ids_per_batch_dtype,
-    const paddle::DataType& kv_num_blocks_dtype,
-    const paddle::DataType& decoder_batch_ids_dtype,
-    const paddle::DataType& decoder_tile_ids_per_batch_dtype,
-    const paddle::DataType& decoder_num_blocks_dtype,
-    const paddle::DataType& set_max_lengths_dtype,
-    const paddle::DataType& max_len_kv_dtype,
-    const paddle::DataType& fmha_out_dtype,
-    const paddle::optional<paddle::DataType>& rotary_embs_dtype,
-    const paddle::optional<paddle::DataType>& attn_mask_dtype,
-    const paddle::optional<paddle::DataType>& qkv_bias_dtype,
-    const paddle::optional<paddle::DataType>& qkv_out_scales_dtype,
-    const paddle::optional<paddle::DataType>& cache_k_quant_scales_dtype,
-    const paddle::optional<paddle::DataType>& cache_v_quant_scales_dtype,
-    const paddle::optional<paddle::DataType>& cache_k_dequant_scales_dtype,
-    const paddle::optional<paddle::DataType>& cache_v_dequant_scales_dtype,
-    const paddle::optional<paddle::DataType>& cache_k_zp_dtype,
-    const paddle::optional<paddle::DataType>& cache_v_zp_dtype,
-    const paddle::optional<paddle::DataType>& out_linear_shifts_dtype,
-    const paddle::optional<paddle::DataType>& out_linear_smooths_dtype,
-    const paddle::optional<paddle::DataType>& mask_offset_dtype,
-    const paddle::optional<paddle::DataType>& kv_signal_data_dtype,
-    const paddle::optional<paddle::DataType>& q_norm_weight_dtype,
-    const paddle::optional<paddle::DataType>& k_norm_weight_dtype,
-    const float rms_norm_eps,
-    const std::string& compute_dtype,
-    const std::string& cache_quant_type_str,
-    const bool use_neox_rotary_style,
-    const bool rope_3d,
-    const int max_input_length,
-    const float quant_max_bound,
-    const float quant_min_bound,
-    const float out_linear_in_scale,
-    const int encoder_block_shape_q,
-    const int decoder_block_shape_q,
-    const int max_partition_size,
-    const int encoder_max_partition_size,
-    const int speculate_max_draft_token_num,
-    const bool causal,
-    const bool speculate_decoder) {
-  return {fmha_out_dtype};
-}
-
-
-
 PD_BUILD_STATIC_OP(append_attention)
     .Inputs({"qkv",
              "key_cache",
@@ -1035,8 +688,8 @@ PD_BUILD_STATIC_OP(append_attention)
              "seq_lens_encoder",
              "seq_lens_decoder",
              "seq_lens_this_time",
-             "batch_id_per_token",
-             "cu_seqlens_q",
+             "padding_offsets",
+             "cum_offsets",
              "block_tables",
              "encoder_batch_ids",
              "encoder_tile_ids_per_batch",
@@ -1061,15 +714,11 @@ PD_BUILD_STATIC_OP(append_attention)
              paddle::Optional("cache_v_zp"),
              paddle::Optional("out_linear_shifts"),
              paddle::Optional("out_linear_smooths"),
-             paddle::Optional("mask_offset"),
-             paddle::Optional("kv_signal_data"),
-             paddle::Optional("q_norm_weight"),
-             paddle::Optional("k_norm_weight")})
-    .Outputs({"fmha_out", "key_cache_out", "value_cache_out"})
+             paddle::Optional("kv_signal_data")})
+    .Outputs({"fmha_out", "qkv_out", "key_cache_out", "value_cache_out"})
     .SetInplaceMap({{"key_cache", "key_cache_out"},
                     {"value_cache", "value_cache_out"}})
-    .Attrs({"rms_norm_eps: float",
-            "compute_type: std::string",
+    .Attrs({"compute_type: std::string",
             "cache_quant_type: std::string",
             "use_neox_rotary_style: bool",
             "rope_3d: bool",
@@ -1083,71 +732,7 @@ PD_BUILD_STATIC_OP(append_attention)
             "encoder_max_partition_size: int",
             "speculate_max_draft_token_num: int",
             "causal: bool",
-            "speculate_decoder: bool",
-            })
+            "speculate_decoder: bool"})
     .SetKernelFn(PD_KERNEL(AppendAttention))
     .SetInferShapeFn(PD_INFER_SHAPE(AppendAttentionInferShape))
     .SetInferDtypeFn(PD_INFER_DTYPE(AppendAttentionInferDtype));
-
-PD_BUILD_STATIC_OP(append_attention_with_output)
-    .Inputs({"qkv",
-             "key_cache",
-             "value_cache",
-             "seq_lens_encoder",
-             "seq_lens_decoder",
-             "seq_lens_this_time",
-             "batch_id_per_token",
-             "cu_seqlens_q",
-             "block_tables",
-             "encoder_batch_ids",
-             "encoder_tile_ids_per_batch",
-             "encoder_num_blocks",
-             "kv_batch_ids",
-             "kv_tile_ids_per_batch",
-             "kv_num_blocks",
-             "decoder_batch_ids",
-             "decoder_tile_ids_per_batch",
-             "decoder_num_blocks",
-             "set_max_lengths",
-             "max_len_kv",
-             "fmha_out",
-             paddle::Optional("rotary_embs"),
-             paddle::Optional("attn_mask"),
-             paddle::Optional("qkv_bias"),
-             paddle::Optional("qkv_out_scales"),
-             paddle::Optional("cache_k_quant_scales"),
-             paddle::Optional("cache_v_quant_scales"),
-             paddle::Optional("cache_k_dequant_scales"),
-             paddle::Optional("cache_v_dequant_scales"),
-             paddle::Optional("cache_k_zp"),
-             paddle::Optional("cache_v_zp"),
-             paddle::Optional("out_linear_shifts"),
-             paddle::Optional("out_linear_smooths"),
-             paddle::Optional("mask_offset"),
-             paddle::Optional("kv_signal_data"),
-             paddle::Optional("q_norm_weight"),
-             paddle::Optional("k_norm_weight")})
-    .Outputs({"fmha_out_out", "qkv_out", "key_cache_out", "value_cache_out"})
-    .SetInplaceMap({{"fmha_out", "fmha_out_out"},
-                    {"key_cache", "key_cache_out"},
-                    {"value_cache", "value_cache_out"}})
-    .Attrs({"rms_norm_eps: float",
-            "compute_type: std::string",
-            "cache_quant_type: std::string",
-            "use_neox_rotary_style: bool",
-            "rope_3d: bool",
-            "max_input_length: int",
-            "quant_max_bound: float",
-            "quant_min_bound: float",
-            "out_linear_in_scale: float",
-            "encoder_block_shape_q: int",
-            "decoder_block_shape_q: int",
-            "max_partition_size: int",
-            "encoder_max_partition_size: int",
-            "speculate_max_draft_token_num: int",
-            "causal: bool",
-            "speculate_decoder: bool",
-            })
-    .SetKernelFn(PD_KERNEL(AppendAttentionWithOutput))
-    .SetInferShapeFn(PD_INFER_SHAPE(AppendAttentionWithOutputInferShape))
-    .SetInferDtypeFn(PD_INFER_DTYPE(AppendAttentionWithOutputInferDtype));

custom_ops/gpu_ops/append_attn/append_attention_c16_impl.cuh

@@ -41,9 +41,8 @@ __global__ void multi_query_append_attention_kernel(
     const int *__restrict__ seq_lens_kv,
     const int *__restrict__ batch_ids,
     const int *__restrict__ tile_ids_per_batch,
-    const int *__restrict__ cu_seqlens_q,
+    const int *__restrict__ cum_offsets,
     const int *__restrict__ block_table,  // [bsz, block_num_per_seq]
-    const int *__restrict__ mask_offset,
     const int max_seq_len,
     const int max_dec_len,
     const int max_block_num_per_seq,
@@ -115,7 +114,8 @@ __global__ void multi_query_append_attention_kernel(
   const uint32_t kv_n_stride = kv_num_heads * BLOCK_SIZE * HEAD_DIM;
   const uint32_t kv_h_stride = BLOCK_SIZE * HEAD_DIM;
   const uint32_t kv_b_stride = HEAD_DIM;
-  const uint32_t q_start_seq_id = cu_seqlens_q[batch_id];
+  const uint32_t q_start_seq_id =
+      batch_id * max_seq_len - __ldg(&cum_offsets[batch_id]);
   const uint32_t q_base_seq_id_this_block =
       (tile_id * NUM_WARPS + wid) * num_frags_x * 16;
   const uint32_t q_offset = q_start_seq_id * q_ori_n_stride +
@@ -142,7 +142,6 @@ __global__ void multi_query_append_attention_kernel(
   } else {
     o_base_ptr_int8 = out + o_offset;
   }
-  const int *mask_offset_this_seq = mask_offset ? mask_offset + q_start_seq_id : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -181,7 +180,7 @@ __global__ void multi_query_append_attention_kernel(
                          kv_len - q_len +
                              tile_id * num_rows_per_block / GROUP_SIZE,
                          chunk_start)))
-              : mask_offset ? 0 : chunk_len) /
+              : chunk_len) /
       (num_frags_z * 16);
   uint32_t k_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
       8 * (tid / 16) + tid % 8, (tid % 16) / 8);
@@ -247,16 +246,12 @@ __global__ void multi_query_append_attention_kernel(
              NUM_WARPS,
              num_frags_x,
              num_frags_y,
-             num_frags_z>(nullptr,
-                          q_base_seq_id_this_block,
+             num_frags_z>(q_base_seq_id_this_block,
                           kv_idx_base,
                           q_len,
                           kv_len,
                           chunk_end,
-                          -1,
-                          s_frag,
-                          mask_offset_this_seq);
-
+                          s_frag);
     }
 
     // update m,d
@@ -410,10 +405,8 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
     const int *__restrict__ seq_lens_kv,
     const int *__restrict__ batch_ids,
     const int *__restrict__ tile_ids_per_batch,
-    const int *__restrict__ cu_seqlens_q,
+    const int *__restrict__ cum_offsets,
     const int *__restrict__ block_table,  // [bsz, block_num_per_seq]
-    const int *__restrict__ mask_offset,
-    const bool *__restrict__ attn_mask,    // [bsz, max_q, max_q] for tree-mask
     const int max_seq_len,
     const int max_dec_len,
     const int max_block_num_per_seq,
@@ -427,8 +420,7 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
     float *__restrict__ tmp_m,      // [token_num, num_chunks, num_heads]
     float *__restrict__ tmp_d,      // [token_num, num_chunks, num_heads]
     OutT *__restrict__ out,
-    const int speculate_max_draft_token_num = 5,
-    const uint32_t attn_mask_len = -1) {
+    const int speculate_max_draft_token_num = 5) {
   constexpr uint32_t num_vecs_per_head = HEAD_DIM / num_elems_per_128b<T>();
   static_assert(NUM_WARP_Q == 1, "NUM_WARP_Q must be 1");
   static_assert(NUM_WARP_KV == 4, "NUM_WARP_KV must be 4");
@@ -485,7 +477,8 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
   const uint32_t kv_n_stride = kv_num_heads * BLOCK_SIZE * HEAD_DIM;
   const uint32_t kv_h_stride = BLOCK_SIZE * HEAD_DIM;
   const uint32_t kv_b_stride = HEAD_DIM;
-  const uint32_t q_start_seq_id = cu_seqlens_q[batch_id];
+  const uint32_t q_start_seq_id =
+      batch_id * max_seq_len - __ldg(&cum_offsets[batch_id]);
   const uint32_t q_base_seq_id_this_block = tile_id * num_frags_x * 16;
   const uint32_t q_offset = q_start_seq_id * q_ori_n_stride +
                             q_head_idx * HEAD_DIM +
@@ -511,7 +504,7 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
           tid % 8 * num_elems_per_128b<T>();
     }
   }
-  const int *mask_offset_this_seq = mask_offset ? mask_offset + q_start_seq_id : nullptr;
+
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -549,9 +542,10 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
   const uint32_t mask_check_iteration =
       (CAUSAL ? (min(chunk_len,
                      sub_if_greater_or_zero(
-                         kv_len - q_len,
+                         kv_len - q_len +
+                             tile_id * num_rows_per_block / GROUP_SIZE,
                          chunk_start)))
-              : mask_offset ? 0 : chunk_len) /
+              : chunk_len) /
       (NUM_WARP_KV * num_frags_z * 16);
 
   uint32_t k_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -619,15 +613,12 @@ __global__ void multi_query_append_attention_warp1_4_kernel(
              NUM_WARPS,
              num_frags_x,
              num_frags_y,
-             num_frags_z>(attn_mask ? attn_mask + batch_id * attn_mask_len *attn_mask_len : nullptr,
-                          q_base_seq_id_this_block,
+             num_frags_z>(q_base_seq_id_this_block,
                           kv_idx_base + wid * num_frags_z * 16,
                           q_len,
                           kv_len,
                           chunk_end,
-                          attn_mask_len,
-                          s_frag,
-                          mask_offset_this_seq);
+                          s_frag);
     }
 
     // update m,d
@@ -784,8 +775,8 @@ void MultiQueryAppendAttention(
     const paddle::Tensor &seq_lens_q,
     const paddle::Tensor &seq_lens_kv,
     const paddle::Tensor &seq_lens_encoder,
-    const paddle::Tensor &batch_id_per_token,
-    const paddle::Tensor &cu_seqlens_q,
+    const paddle::Tensor &padding_offsets,
+    const paddle::Tensor &cum_offsets,
     const paddle::Tensor &block_table,
     const paddle::Tensor &batch_ids,
     const paddle::Tensor &tile_ids_per_batch,
@@ -891,9 +882,8 @@ void MultiQueryAppendAttention(
           seq_lens_kv.data<int>(),
           batch_ids.data<int>(),
           tile_ids_per_batch.data<int>(),
-          cu_seqlens_q.data<int>(),
+          cum_offsets.data<int>(),
           block_table.data<int>(),
-          meta_data.mask_offset,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -949,9 +939,8 @@ void MultiQueryAppendAttention(
           seq_lens_kv.data<int>(),
           batch_ids.data<int>(),
           tile_ids_per_batch.data<int>(),
-          cu_seqlens_q.data<int>(),
+          cum_offsets.data<int>(),
           block_table.data<int>(),
-          meta_data.mask_offset,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -985,7 +974,7 @@ void MultiQueryAppendAttention(
                 seq_lens_q.data<int>(),
                 seq_lens_kv.data<int>(),
                 seq_lens_encoder.data<int>(),
-                cu_seqlens_q.data<int>(),
+                cum_offsets.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
                                  const_cast<T *>(shift_bias.get().data<T>()))
                            : nullptr,
@@ -1020,8 +1009,7 @@ void MultiQueryAppendAttention(
                 seq_lens_q.data<int>(),
                 seq_lens_kv.data<int>(),
                 seq_lens_encoder.data<int>(),
-                batch_id_per_token.data<int>(),
-                cu_seqlens_q.data<int>(),
+                padding_offsets.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
                                  const_cast<T *>(shift_bias.get().data<T>()))
                            : nullptr,
@@ -1074,18 +1062,12 @@ void MultiQueryAppendAttention(
     if (!is_decoder) {
       chunk_size = static_cast<uint32_t>(encoder_max_partition_size);
     }
+    const int num_chunks = div_up(max_dec_len, chunk_size);
 
-    uint32_t attn_mask_len;
-    if (attn_mask) {
-        attn_mask_len = attn_mask.get().shape()[1];
-    } else {
-        attn_mask_len = -1;
-    }
-
-    const int num_chunks = div_up(max_seq_len, chunk_size);
     dim3 grids(num_blocks_x_cpu, num_chunks, kv_num_heads);
     dim3 blocks(32, num_warps);
-    if (num_chunks <= 0) {
+
+    if (num_chunks <= 1) {
       auto nosplit_kv_kernel =
           multi_query_append_attention_warp1_4_kernel<NV_TYPE,
                                                       false,
@@ -1121,11 +1103,8 @@ void MultiQueryAppendAttention(
           seq_lens_kv.data<int>(),
           batch_ids.data<int>(),
           tile_ids_per_batch.data<int>(),
-          cu_seqlens_q.data<int>(),
+          cum_offsets.data<int>(),
           block_table.data<int>(),
-          meta_data.mask_offset,
-          attn_mask ? const_cast<bool *>(attn_mask.get().data<bool>())
-                        : nullptr,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1138,8 +1117,7 @@ void MultiQueryAppendAttention(
           nullptr,
           nullptr,
           reinterpret_cast<OUT_NV_TYPE *>(out->data<OutT>()),
-          speculate_max_draft_token_num,
-          attn_mask_len);
+          speculate_max_draft_token_num);
     } else {
       phi::Allocator::AllocationPtr tmp_workspace, tmp_m, tmp_d;
       if (is_decoder) {
@@ -1184,8 +1162,8 @@ void MultiQueryAppendAttention(
           reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_k.data<T>())),
           reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_v.data<T>())),
           shift_bias ? reinterpret_cast<NV_TYPE *>(
-                            const_cast<T *>(shift_bias.get().data<T>()))
-                      : nullptr,
+                           const_cast<T *>(shift_bias.get().data<T>()))
+                     : nullptr,
           smooth_weight ? reinterpret_cast<NV_TYPE *>(
                               const_cast<T *>(smooth_weight.get().data<T>()))
                         : nullptr,
@@ -1193,11 +1171,8 @@ void MultiQueryAppendAttention(
           seq_lens_kv.data<int>(),
           batch_ids.data<int>(),
           tile_ids_per_batch.data<int>(),
-          cu_seqlens_q.data<int>(),
+          cum_offsets.data<int>(),
           block_table.data<int>(),
-          meta_data.mask_offset,
-          attn_mask ? const_cast<bool *>(attn_mask.get().data<bool>())
-                        : nullptr,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1210,8 +1185,7 @@ void MultiQueryAppendAttention(
           static_cast<float *>(tmp_m->ptr()),
           static_cast<float *>(tmp_d->ptr()),
           reinterpret_cast<OUT_NV_TYPE *>(out->data<OutT>()),
-          speculate_max_draft_token_num,
-          attn_mask_len);
+          speculate_max_draft_token_num);
 
       // merge
       constexpr int vec_size = num_elems_per_128b<NV_TYPE>();
@@ -1233,10 +1207,10 @@ void MultiQueryAppendAttention(
                 seq_lens_q.data<int>(),
                 seq_lens_kv.data<int>(),
                 seq_lens_encoder.data<int>(),
-                cu_seqlens_q.data<int>(),
+                cum_offsets.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
-                                  const_cast<T *>(shift_bias.get().data<T>()))
-                            : nullptr,
+                                 const_cast<T *>(shift_bias.get().data<T>()))
+                           : nullptr,
                 smooth_weight ? reinterpret_cast<NV_TYPE *>(const_cast<T *>(
                                     smooth_weight.get().data<T>()))
                               : nullptr,
@@ -1253,14 +1227,14 @@ void MultiQueryAppendAttention(
         constexpr int blockx = HEAD_DIM / vec_size;
         constexpr int blocky = (128 + blockx - 1) / blockx;
         dim3 grids_merge(min(sm_count * 4, token_num),
-                          num_heads);
+                         num_heads);
         dim3 blocks_merge(blockx, blocky);
         merge_multi_chunks_v2_kernel<NV_TYPE,
-                                      vec_size,
-                                      blocky,
-                                      HEAD_DIM,
-                                      OUT_NV_TYPE,
-                                      ENABLE_PREFILL>
+                                     vec_size,
+                                     blocky,
+                                     HEAD_DIM,
+                                     OUT_NV_TYPE,
+                                     ENABLE_PREFILL>
             <<<grids_merge, blocks_merge, 0, stream>>>(
                 reinterpret_cast<NV_TYPE *>(tmp_workspace->ptr()),
                 static_cast<float *>(tmp_m->ptr()),
@@ -1268,11 +1242,10 @@ void MultiQueryAppendAttention(
                 seq_lens_q.data<int>(),
                 seq_lens_kv.data<int>(),
                 seq_lens_encoder.data<int>(),
-                batch_id_per_token.data<int>(),
-                cu_seqlens_q.data<int>(),
+                padding_offsets.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
-                                  const_cast<T *>(shift_bias.get().data<T>()))
-                            : nullptr,
+                                 const_cast<T *>(shift_bias.get().data<T>()))
+                           : nullptr,
                 smooth_weight ? reinterpret_cast<NV_TYPE *>(const_cast<T *>(
                                     smooth_weight.get().data<T>()))
                               : nullptr,
@@ -1316,8 +1289,8 @@ void CascadeAppendAttentionC16Kernel(
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,
@@ -1379,8 +1352,8 @@ void CascadeAppendAttentionC16Kernel(
                                 seq_lens_q,
                                 seq_lens_kv,
                                 seq_lens_encoder,
-                                batch_id_per_token,
-                                cu_seqlens_q,
+                                padding_offsets,
+                                cum_offsets,
                                 block_table,
                                 batch_ids,
                                 tile_ids_per_batch,

custom_ops/gpu_ops/append_attn/append_attention_c4_impl.cuh

@@ -46,9 +46,8 @@ __global__ void multi_query_append_attention_c4_kernel(
     const int *__restrict__ seq_lens_kv,
     const int *__restrict__ batch_ids,
     const int *__restrict__ tile_ids_per_batch,
-    const int *__restrict__ cu_seqlens_q,
+    const int *__restrict__ cum_offsets,
     const int *__restrict__ block_table,  // [bsz, block_num_per_seq]
-    const int *__restrict__ mask_offset,
     const int max_seq_len,
     const int max_dec_len,
     const int max_block_num_per_seq,
@@ -145,7 +144,8 @@ __global__ void multi_query_append_attention_c4_kernel(
   const uint32_t kv_h_stride = BLOCK_SIZE * HEAD_DIM / 2;
   const uint32_t kv_b_stride = HEAD_DIM / 2;
   const uint32_t kv_d_stride = BLOCK_SIZE / 2;
-  const uint32_t q_start_seq_id = cu_seqlens_q[batch_id];
+  const uint32_t q_start_seq_id =
+      batch_id * max_seq_len - __ldg(&cum_offsets[batch_id]);
   const uint32_t q_base_seq_id_this_block =
       (tile_id * NUM_WARPS + wid) * num_frags_x * 16;
   const uint32_t q_offset = q_start_seq_id * q_ori_n_stride +
@@ -173,7 +173,6 @@ __global__ void multi_query_append_attention_c4_kernel(
   } else {
     o_base_ptr_int8 = out + o_offset;
   }
-  const int *mask_offset_this_seq = mask_offset ? mask_offset + q_start_seq_id : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -250,7 +249,7 @@ __global__ void multi_query_append_attention_c4_kernel(
                          kv_len - q_len +
                              tile_id * num_rows_per_block / GROUP_SIZE,
                          chunk_start)))
-              : mask_offset ? 0 : chunk_len) /
+              : chunk_len) /
       (num_frags_z * 16);
 
   uint32_t k_smem_offset_r =
@@ -335,15 +334,12 @@ __global__ void multi_query_append_attention_c4_kernel(
              NUM_WARPS,
              num_frags_x,
              num_frags_y,
-             num_frags_z>(nullptr,
-                          q_base_seq_id_this_block,
+             num_frags_z>(q_base_seq_id_this_block,
                           kv_idx_base,
                           q_len,
                           kv_len,
                           chunk_end,
-                          -1,
-                          s_frag,
-                          mask_offset_this_seq);
+                          s_frag);
     }
 
     update_mdo_states<num_frags_x, num_frags_y, num_frags_z>(
@@ -508,10 +504,8 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
     const int *__restrict__ seq_lens_kv,
     const int *__restrict__ batch_ids,
     const int *__restrict__ tile_ids_per_batch,
-    const int *__restrict__ cu_seqlens_q,
+    const int *__restrict__ cum_offsets,
     const int *__restrict__ block_table,  // [bsz, block_num_per_seq]
-    const int *__restrict__ mask_offset,
-    const bool *__restrict__ attn_mask,    // [bsz, max_q, max_q] for tree-mask
     const int max_seq_len,
     const int max_dec_len,
     const int max_block_num_per_seq,
@@ -525,8 +519,7 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
     float *__restrict__ tmp_m,      // [token_num, num_chunks, num_heads]
     float *__restrict__ tmp_d,      // [token_num, num_chunks, num_heads]
     OutT *__restrict__ out,
-    const int speculate_max_draft_token_num = 5,
-    const uint32_t attn_mask_len = -1) {
+    const int speculate_max_draft_token_num = 5) {
   constexpr uint32_t num_vecs_per_head = HEAD_DIM / num_elems_per_128b<T>();
   constexpr uint32_t num_vecs_per_head_k =
       HEAD_DIM / 2 / num_elems_per_128b<CacheT>();
@@ -608,7 +601,8 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
   const uint32_t kv_h_stride = BLOCK_SIZE * HEAD_DIM / 2;
   const uint32_t kv_b_stride = HEAD_DIM / 2;
   const uint32_t kv_d_stride = BLOCK_SIZE / 2;
-  const uint32_t q_start_seq_id = cu_seqlens_q[batch_id];
+  const uint32_t q_start_seq_id =
+      batch_id * max_seq_len - __ldg(&cum_offsets[batch_id]);
   const uint32_t q_base_seq_id_this_block = tile_id * num_frags_x * 16;
   const uint32_t q_offset = q_start_seq_id * q_ori_n_stride +
                             q_head_idx * HEAD_DIM +
@@ -635,7 +629,7 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
           tid % 8 * num_elems_per_128b<T>();
     }
   }
-  const int *mask_offset_this_seq = mask_offset ? mask_offset + q_start_seq_id : nullptr;
+
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -711,9 +705,10 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
   const uint32_t mask_check_iteration =
       (CAUSAL ? (min(chunk_len,
                      sub_if_greater_or_zero(
-                         kv_len - q_len,
+                         kv_len - q_len +
+                             tile_id * num_rows_per_block / GROUP_SIZE,
                          chunk_start)))
-              : mask_offset ? 0 : chunk_len) /
+              : chunk_len) /
       (NUM_WARP_KV * num_frags_z * 16);
 
   uint32_t k_smem_offset_r =
@@ -795,15 +790,12 @@ __global__ void multi_query_append_attention_c4_warp1_4_kernel(
              NUM_WARPS,
              num_frags_x,
              num_frags_y,
-             num_frags_z>(attn_mask ? attn_mask + batch_id * attn_mask_len *attn_mask_len : nullptr,
-                          q_base_seq_id_this_block,
+             num_frags_z>(q_base_seq_id_this_block,
                           kv_idx_base + wid * num_frags_z * 16,
                           q_len,
                           kv_len,
                           chunk_end,
-                          attn_mask_len,
-                          s_frag,
-                          mask_offset_this_seq);
+                          s_frag);
     }
 
     update_mdo_states<num_frags_x, num_frags_y, num_frags_z>(
@@ -970,8 +962,8 @@ void MultiQueryAppendC4Attention(
     const paddle::Tensor &seq_lens_q,
     const paddle::Tensor &seq_lens_kv,
     const paddle::Tensor &seq_lens_encoder,
-    const paddle::Tensor &batch_id_per_token,
-    const paddle::Tensor &cu_seqlens_q,
+    const paddle::Tensor &padding_offsets,
+    const paddle::Tensor &cum_offsets,
     const paddle::Tensor &block_table,
     const paddle::Tensor &batch_ids,
     const paddle::Tensor &tile_ids_per_batch,
@@ -1096,9 +1088,8 @@ void MultiQueryAppendC4Attention(
           seq_lens_kv.data<int>(),
           batch_ids.data<int>(),
           tile_ids_per_batch.data<int>(),
-          cu_seqlens_q.data<int>(),
+          cum_offsets.data<int>(),
           block_table.data<int>(),
-          meta_data.mask_offset,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1160,9 +1151,8 @@ void MultiQueryAppendC4Attention(
           seq_lens_kv.data<int>(),
           batch_ids.data<int>(),
           tile_ids_per_batch.data<int>(),
-          cu_seqlens_q.data<int>(),
+          cum_offsets.data<int>(),
           block_table.data<int>(),
-          meta_data.mask_offset,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1196,7 +1186,7 @@ void MultiQueryAppendC4Attention(
                 seq_lens_q.data<int>(),
                 seq_lens_kv.data<int>(),
                 seq_lens_encoder.data<int>(),
-                cu_seqlens_q.data<int>(),
+                cum_offsets.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
                                  const_cast<T *>(shift_bias.get().data<T>()))
                            : nullptr,
@@ -1231,8 +1221,7 @@ void MultiQueryAppendC4Attention(
                 seq_lens_q.data<int>(),
                 seq_lens_kv.data<int>(),
                 seq_lens_encoder.data<int>(),
-                batch_id_per_token.data<int>(),
-                cu_seqlens_q.data<int>(),
+                padding_offsets.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
                                  const_cast<T *>(shift_bias.get().data<T>()))
                            : nullptr,
@@ -1297,18 +1286,10 @@ void MultiQueryAppendC4Attention(
     if (!is_decoder) {
       chunk_size = static_cast<uint32_t>(encoder_max_partition_size);
     }
-
-    const int num_chunks = div_up(max_seq_len, chunk_size);
-    uint32_t attn_mask_len;
-    if (attn_mask) {
-        attn_mask_len = attn_mask.get().shape()[1];
-    } else {
-        attn_mask_len = -1;
-    }
-
+    const int num_chunks = div_up(max_dec_len, chunk_size);
     dim3 grids(num_blocks_x_cpu, num_chunks, kv_num_heads);
     dim3 blocks(32, num_warps);
-    if (num_chunks <= 0) {
+    if (num_chunks <= 1) {
       auto nosplit_kv_kernel =
           multi_query_append_attention_c4_warp1_4_kernel<NV_TYPE,
                                                          uint8_t,
@@ -1352,11 +1333,8 @@ void MultiQueryAppendC4Attention(
           seq_lens_kv.data<int>(),
           batch_ids.data<int>(),
           tile_ids_per_batch.data<int>(),
-          cu_seqlens_q.data<int>(),
+          cum_offsets.data<int>(),
           block_table.data<int>(),
-          meta_data.mask_offset,
-          attn_mask ? const_cast<bool *>(attn_mask.get().data<bool>())
-                        : nullptr,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1369,8 +1347,7 @@ void MultiQueryAppendC4Attention(
           nullptr,
           nullptr,
           reinterpret_cast<OUT_NV_TYPE *>(out->data<OutT>()),
-          speculate_max_draft_token_num,
-          attn_mask_len);
+          speculate_max_draft_token_num);
     } else {
       phi::Allocator::AllocationPtr tmp_workspace, tmp_m, tmp_d;
       if (is_decoder) {
@@ -1416,15 +1393,15 @@ void MultiQueryAppendC4Attention(
           const_cast<uint8_t *>(cache_v.data<uint8_t>()),
           reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_k_scale.data<T>())),
           cache_k_zp ? reinterpret_cast<NV_TYPE *>(
-                            const_cast<T *>(cache_k_zp.get().data<T>()))
-                      : nullptr,
+                           const_cast<T *>(cache_k_zp.get().data<T>()))
+                     : nullptr,
           reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_v_scale.data<T>())),
           cache_v_zp ? reinterpret_cast<NV_TYPE *>(
-                            const_cast<T *>(cache_v_zp.get().data<T>()))
-                      : nullptr,
+                           const_cast<T *>(cache_v_zp.get().data<T>()))
+                     : nullptr,
           shift_bias ? reinterpret_cast<NV_TYPE *>(
-                            const_cast<T *>(shift_bias.get().data<T>()))
-                      : nullptr,
+                           const_cast<T *>(shift_bias.get().data<T>()))
+                     : nullptr,
           smooth_weight ? reinterpret_cast<NV_TYPE *>(
                               const_cast<T *>(smooth_weight.get().data<T>()))
                         : nullptr,
@@ -1432,11 +1409,8 @@ void MultiQueryAppendC4Attention(
           seq_lens_kv.data<int>(),
           batch_ids.data<int>(),
           tile_ids_per_batch.data<int>(),
-          cu_seqlens_q.data<int>(),
+          cum_offsets.data<int>(),
           block_table.data<int>(),
-          meta_data.mask_offset,
-          attn_mask ? const_cast<bool *>(attn_mask.get().data<bool>())
-                        : nullptr,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1449,8 +1423,7 @@ void MultiQueryAppendC4Attention(
           static_cast<float *>(tmp_m->ptr()),
           static_cast<float *>(tmp_d->ptr()),
           reinterpret_cast<OUT_NV_TYPE *>(out->data<OutT>()),
-          speculate_max_draft_token_num,
-          attn_mask_len);
+          speculate_max_draft_token_num);
       // merge
       constexpr int vec_size = num_elems_per_128b<NV_TYPE>();
       if (is_decoder) {
@@ -1471,10 +1444,10 @@ void MultiQueryAppendC4Attention(
                 seq_lens_q.data<int>(),
                 seq_lens_kv.data<int>(),
                 seq_lens_encoder.data<int>(),
-                cu_seqlens_q.data<int>(),
+                cum_offsets.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
-                                  const_cast<T *>(shift_bias.get().data<T>()))
-                            : nullptr,
+                                 const_cast<T *>(shift_bias.get().data<T>()))
+                           : nullptr,
                 smooth_weight ? reinterpret_cast<NV_TYPE *>(const_cast<T *>(
                                     smooth_weight.get().data<T>()))
                               : nullptr,
@@ -1491,14 +1464,14 @@ void MultiQueryAppendC4Attention(
         constexpr int blockx = HEAD_DIM / vec_size;
         constexpr int blocky = (128 + blockx - 1) / blockx;
         dim3 grids_merge(min(sm_count * 4, token_num),
-                          num_heads);
+                         num_heads);
         dim3 blocks_merge(blockx, blocky);
         merge_multi_chunks_v2_kernel<NV_TYPE,
-                                      vec_size,
-                                      blocky,
-                                      HEAD_DIM,
-                                      OUT_NV_TYPE,
-                                      ENABLE_PREFILL>
+                                     vec_size,
+                                     blocky,
+                                     HEAD_DIM,
+                                     OUT_NV_TYPE,
+                                     ENABLE_PREFILL>
             <<<grids_merge, blocks_merge, 0, stream>>>(
                 reinterpret_cast<NV_TYPE *>(tmp_workspace->ptr()),
                 static_cast<float *>(tmp_m->ptr()),
@@ -1506,11 +1479,10 @@ void MultiQueryAppendC4Attention(
                 seq_lens_q.data<int>(),
                 seq_lens_kv.data<int>(),
                 seq_lens_encoder.data<int>(),
-                batch_id_per_token.data<int>(),
-                cu_seqlens_q.data<int>(),
+                padding_offsets.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
-                                  const_cast<T *>(shift_bias.get().data<T>()))
-                            : nullptr,
+                                 const_cast<T *>(shift_bias.get().data<T>()))
+                           : nullptr,
                 smooth_weight ? reinterpret_cast<NV_TYPE *>(const_cast<T *>(
                                     smooth_weight.get().data<T>()))
                               : nullptr,
@@ -1554,8 +1526,8 @@ void CascadeAppendAttentionC4Kernel(
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,
@@ -1621,8 +1593,8 @@ void CascadeAppendAttentionC4Kernel(
                                 seq_lens_q,
                                 seq_lens_kv,
                                 seq_lens_encoder,
-                                batch_id_per_token,
-                                cu_seqlens_q,
+                                padding_offsets,
+                                cum_offsets,
                                 block_table,
                                 batch_ids,
                                 tile_ids_per_batch,

custom_ops/gpu_ops/append_attn/append_attention_c8_impl.cuh

@@ -32,24 +32,22 @@ template <typename T,
           typename OutT = T,
           bool ENABLE_PREFILL = true,
           bool is_scale_channel_wise = false,
-          bool IsFP8 = false,
-          bool IsDynamicC8 = false>
+          bool IsFP8=false>
 __global__ void multi_query_append_attention_c8_kernel(
     T *__restrict__ q,             // [token_num, (num_heads + 2* kv_num_head) * head_dim]
     CacheT *__restrict__ cache_k,  // [max_block_num, num_heads, block_size,
                                    // head_dim]
     CacheT *__restrict__ cache_v,
-    const T *__restrict__ cache_k_scale,  // [num_kv_heads] or [max_block_num, num_heads, block_size]
-    const T *__restrict__ cache_v_scale,  // [num_kv_heads] or [max_block_num, num_heads, block_size]
+    const T *__restrict__ cache_k_scale,  // [num_kv_heads]
+    const T *__restrict__ cache_v_scale,  // [num_kv_heads]
     const T *__restrict__ shift_bias,     // [q_num_heads * HEAD_DIM]
     const T *__restrict__ smooth_weight,  // [q_num_heads * HEAD_DIM]
     const int *__restrict__ seq_lens,
     const int *__restrict__ seq_lens_kv,
     const int *__restrict__ batch_ids,
     const int *__restrict__ tile_ids_per_batch,
-    const int *__restrict__ cu_seqlens_q,
+    const int *__restrict__ cum_offsets,
     const int *__restrict__ block_table,  // [bsz, block_num_per_seq]
-    const int *__restrict__ mask_offset,
     const int max_seq_len,
     const int max_dec_len,
     const int max_block_num_per_seq,
@@ -92,30 +90,28 @@ __global__ void multi_query_append_attention_c8_kernel(
     return;
   }
 
-  T cache_k_scale_reg[IsDynamicC8 ? num_frags_z * 2 : num_frags_y * 4];
-  T cache_v_scale_reg[IsDynamicC8 ? num_frags_z * 4 : num_frags_y * 2];
-  if constexpr (!IsDynamicC8) {
-    if constexpr (is_scale_channel_wise) {
-      int scale_col_base = threadIdx.x % 4 * 2 + kv_head_idx * HEAD_DIM;
-      const T *cache_k_scale_cur_head = cache_k_scale + scale_col_base;
-      for (int i = 0; i < num_frags_y; ++i) {
-        const int scale_idx = i * 16;
-        cache_k_scale_reg[i * 4] = cache_k_scale_cur_head[scale_idx];
-        cache_k_scale_reg[i * 4 + 1] = cache_k_scale_cur_head[scale_idx + 1];
-        cache_k_scale_reg[i * 4 + 2] = cache_k_scale_cur_head[scale_idx + 8];
-        cache_k_scale_reg[i * 4 + 3] = cache_k_scale_cur_head[scale_idx + 9];
-      }
-      scale_col_base = threadIdx.x / 4 + kv_head_idx * HEAD_DIM;
-      const T *cache_v_scale_cur_head = cache_v_scale + scale_col_base;
-      for (int i = 0; i < num_frags_y; ++i) {
-        const int scale_idx = i * 16;
-        cache_v_scale_reg[i * 2] = cache_v_scale_cur_head[scale_idx];
-        cache_v_scale_reg[i * 2 + 1] = cache_v_scale_cur_head[scale_idx + 8];
-      }
-    } else {
-      cache_k_scale_reg[0] = cache_k_scale[kv_head_idx];
-      cache_v_scale_reg[0] = cache_v_scale[kv_head_idx];
+  T cache_k_scale_reg[num_frags_y * 4];
+  T cache_v_scale_reg[num_frags_y * 2];
+  if (is_scale_channel_wise) {
+    int scale_col_base = threadIdx.x % 4 * 2 + kv_head_idx * HEAD_DIM;
+    const T *cache_k_scale_cur_head = cache_k_scale + scale_col_base;
+    for (int i = 0; i < num_frags_y; ++i) {
+      const int scale_idx = i * 16;
+      cache_k_scale_reg[i * 4] = cache_k_scale_cur_head[scale_idx];
+      cache_k_scale_reg[i * 4 + 1] = cache_k_scale_cur_head[scale_idx + 1];
+      cache_k_scale_reg[i * 4 + 2] = cache_k_scale_cur_head[scale_idx + 8];
+      cache_k_scale_reg[i * 4 + 3] = cache_k_scale_cur_head[scale_idx + 9];
     }
+    scale_col_base = threadIdx.x / 4 + kv_head_idx * HEAD_DIM;
+    const T *cache_v_scale_cur_head = cache_v_scale + scale_col_base;
+    for (int i = 0; i < num_frags_y; ++i) {
+      const int scale_idx = i * 16;
+      cache_v_scale_reg[i * 2] = cache_v_scale_cur_head[scale_idx];
+      cache_v_scale_reg[i * 2 + 1] = cache_v_scale_cur_head[scale_idx + 8];
+    }
+  } else {
+    cache_k_scale_reg[0] = cache_k_scale[kv_head_idx];
+    cache_v_scale_reg[0] = cache_v_scale[kv_head_idx];
   }
 
   const uint32_t q_end =
@@ -155,7 +151,8 @@ __global__ void multi_query_append_attention_c8_kernel(
   const uint32_t kv_h_stride = BLOCK_SIZE * HEAD_DIM;
   const uint32_t kv_b_stride = HEAD_DIM;
   const uint32_t kv_d_stride = BLOCK_SIZE;
-  const uint32_t q_start_seq_id = cu_seqlens_q[batch_id];
+  const uint32_t q_start_seq_id =
+      batch_id * max_seq_len - __ldg(&cum_offsets[batch_id]);
   const uint32_t q_base_seq_id_this_block =
       (tile_id * NUM_WARPS + wid) * num_frags_x * 16;
   const uint32_t q_offset = q_start_seq_id * q_ori_n_stride +
@@ -183,7 +180,6 @@ __global__ void multi_query_append_attention_c8_kernel(
   } else {
     o_base_ptr_int8 = out + o_offset;
   }
-  const int *mask_offset_this_seq = mask_offset ? mask_offset + q_start_seq_id : nullptr;
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -204,17 +200,6 @@ __global__ void multi_query_append_attention_c8_kernel(
   smem_t k_smem(smem + NUM_WARPS * num_frags_x * 16 * HEAD_DIM * sizeof(T)),
       v_smem(smem + NUM_WARPS * num_frags_x * 16 * HEAD_DIM * sizeof(T) +
              num_frags_z * 16 * HEAD_DIM * sizeof(CacheT));
-  T* k_smem_scale_ptr = nullptr;
-  T* v_smem_scale_ptr = nullptr;
-  smem_t k_scale_smem;
-  smem_t v_scale_smem;
-  if constexpr (IsDynamicC8) {
-    k_smem_scale_ptr = reinterpret_cast<T*>(smem + NUM_WARPS * num_frags_x * 16 * HEAD_DIM * sizeof(T) +
-                                         num_frags_z * 16 * HEAD_DIM * sizeof(CacheT) * 2);
-    v_smem_scale_ptr = k_smem_scale_ptr + num_frags_z * 16;
-    k_scale_smem.base = reinterpret_cast<b128_t*>(k_smem_scale_ptr);
-    v_scale_smem.base = reinterpret_cast<b128_t*>(v_smem_scale_ptr);
-  }
 
 
   const uint32_t num_iterations = div_up(
@@ -232,7 +217,7 @@ __global__ void multi_query_append_attention_c8_kernel(
                          kv_len - q_len +
                              tile_id * num_rows_per_block / GROUP_SIZE,
                          chunk_start)))
-              : mask_offset ? 0 : chunk_len) /
+              : chunk_len) /
       (num_frags_z * 16);
 
   uint32_t k_smem_offset_r =
@@ -275,20 +260,6 @@ __global__ void multi_query_append_attention_c8_kernel(
                                      kv_idx_base,
                                      chunk_end,
                                      const_k_offset);
-  if constexpr (IsDynamicC8) {
-    produce_kv_dynamic_scale_gmem2smem_async<SharedMemFillMode::kFillZero,
-                                             BLOCK_SIZE,
-                                             num_frags_z,
-                                             NUM_WARP_Q>(
-      k_scale_smem,
-      block_table_now,
-      cache_k_scale,
-      kv_idx_base,
-      kv_num_heads,
-      kv_head_idx,
-      chunk_end
-    );
-  }
   commit_group();
   produce_v_blockwise_c8<SharedMemFillMode::kNoFill,
                          NUM_WARPS,
@@ -306,34 +277,14 @@ __global__ void multi_query_append_attention_c8_kernel(
                                      kv_idx_base,
                                      chunk_end,
                                      const_v_offset);
-  if constexpr (IsDynamicC8) {
-    produce_kv_dynamic_scale_gmem2smem_async<SharedMemFillMode::kFillZero,
-                                             BLOCK_SIZE,
-                                             num_frags_z,
-                                             NUM_WARP_Q>(
-      v_scale_smem,
-      block_table_now,
-      cache_v_scale,
-      kv_idx_base,
-      kv_num_heads,
-      kv_head_idx,
-      chunk_end
-    );
-  }
   commit_group();
 
 #pragma unroll 1
   for (uint32_t iter = 0; iter < num_iterations; ++iter) {
     wait_group<1>();
     __syncthreads();
-    if constexpr (IsDynamicC8) {
-      produce_k_dynamic_scale_smem2reg<BLOCK_SIZE, num_frags_z, NUM_WARP_Q, T>(
-        k_smem_scale_ptr,
-        cache_k_scale_reg
-      );
-    }
     // s = qk
-    compute_qk_c8<num_frags_x, num_frags_y, num_frags_z, T, CacheT, is_scale_channel_wise, IsFP8, IsDynamicC8>(
+    compute_qk_c8<num_frags_x, num_frags_y, num_frags_z, T, CacheT, is_scale_channel_wise, IsFP8>(
         &qo_smem,
         &q_smem_offset_r,
         &k_smem,
@@ -350,15 +301,12 @@ __global__ void multi_query_append_attention_c8_kernel(
              NUM_WARPS,
              num_frags_x,
              num_frags_y,
-             num_frags_z>(nullptr,
-                          q_base_seq_id_this_block,
+             num_frags_z>(q_base_seq_id_this_block,
                           kv_idx_base,
                           q_len,
                           kv_len,
                           chunk_end,
-                          -1,
-                          s_frag,
-                          mask_offset_this_seq);
+                          s_frag);
     }
 
     // update m,d
@@ -366,7 +314,6 @@ __global__ void multi_query_append_attention_c8_kernel(
         s_frag, o_frag, m_frag, d_frag);
     __syncthreads();
 
-    const int ori_kv_idx_base = kv_idx_base;
     kv_idx_base += num_frags_z * 16;
     produce_k_blockwise_c8<SharedMemFillMode::kNoFill,
                            NUM_WARPS,
@@ -384,29 +331,9 @@ __global__ void multi_query_append_attention_c8_kernel(
                                        kv_idx_base,
                                        chunk_end,
                                        const_k_offset);
-    if constexpr (IsDynamicC8) {
-      produce_kv_dynamic_scale_gmem2smem_async<SharedMemFillMode::kFillZero,
-                                              BLOCK_SIZE,
-                                              num_frags_z,
-                                              NUM_WARP_Q>(
-        k_scale_smem,
-        block_table_now,
-        cache_k_scale,
-        kv_idx_base,
-        kv_num_heads,
-        kv_head_idx,
-        chunk_end
-      );
-    }
     commit_group();
     wait_group<1>();
     __syncthreads();
-    if constexpr (IsDynamicC8) {
-      produce_v_dynamic_scale_smem2reg<BLOCK_SIZE, num_frags_z, NUM_WARP_Q, T>(
-        v_smem_scale_ptr,
-        cache_v_scale_reg
-      );
-    }
 
     // compute sfm*v
     compute_sfm_v_c8<num_frags_x,
@@ -415,9 +342,7 @@ __global__ void multi_query_append_attention_c8_kernel(
                      BLOCK_SIZE,
                      T,
                      CacheT,
-                     is_scale_channel_wise,
-                     IsFP8,
-                     IsDynamicC8>(
+                     is_scale_channel_wise, IsFP8>(
         &v_smem, &v_smem_offset_r, s_frag, o_frag, d_frag, cache_v_scale_reg);
     __syncthreads();
 
@@ -437,20 +362,6 @@ __global__ void multi_query_append_attention_c8_kernel(
                                        kv_idx_base,
                                        chunk_end,
                                        const_v_offset);
-    if constexpr (IsDynamicC8) {
-      produce_kv_dynamic_scale_gmem2smem_async<SharedMemFillMode::kFillZero,
-                                              BLOCK_SIZE,
-                                              num_frags_z,
-                                              NUM_WARP_Q>(
-        v_scale_smem,
-        block_table_now,
-        cache_v_scale,
-        kv_idx_base,
-        kv_num_heads,
-        kv_head_idx,
-        chunk_end
-      );
-    }
     commit_group();
 
   }
@@ -548,25 +459,22 @@ template <typename T,
           typename OutT = T,
           bool ENABLE_PREFILL = true,
           bool is_scale_channel_wise=false,
-          bool IsFP8 = false,
-          bool IsDynamicC8 = false>
+          bool IsFP8=false>
 __global__ void multi_query_append_attention_c8_warp1_4_kernel(
     T *__restrict__ q,             // [token_num, (num_heads + 2* kv_num_head) * head_dim]
     CacheT *__restrict__ cache_k,  // [max_block_num, num_heads, block_size,
                                    // head_dim]
     CacheT *__restrict__ cache_v,
-    const T *__restrict__ cache_k_scale,  // [num_kv_heads] or [max_block_num, num_heads, block_size]
-    const T *__restrict__ cache_v_scale,  // [num_kv_heads] or [max_block_num, num_heads, block_size]
+    const T *__restrict__ cache_k_scale,  // [num_kv_heads, head_dim]
+    const T *__restrict__ cache_v_scale,  // [num_kv_heads, head_dim]
     const T *__restrict__ shift_bias,     // [q_num_heads * HEAD_DIM]
     const T *__restrict__ smooth_weight,  // [q_num_heads * HEAD_DIM]
     const int *__restrict__ seq_lens,
     const int *__restrict__ seq_lens_kv,
     const int *__restrict__ batch_ids,
     const int *__restrict__ tile_ids_per_batch,
-    const int *__restrict__ cu_seqlens_q,
+    const int *__restrict__ cum_offsets,
     const int *__restrict__ block_table,  // [bsz, block_num_per_seq]
-    const int *__restrict__ mask_offset,
-    const bool *__restrict__ attn_mask,   // [bsz, max_q, max_q] for tree-mask
     const int max_seq_len,
     const int max_dec_len,
     const int max_block_num_per_seq,
@@ -580,8 +488,7 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
     float *__restrict__ tmp_m,      // [token_num, num_chunks, num_heads]
     float *__restrict__ tmp_d,      // [token_num, num_chunks, num_heads]
     OutT *__restrict__ out,
-    const int speculate_max_draft_token_num = 5,
-    const uint32_t attn_mask_len = -1) {
+    const int speculate_max_draft_token_num = 5) {
   constexpr uint32_t num_vecs_per_head = HEAD_DIM / num_elems_per_128b<T>();
   constexpr uint32_t num_vecs_per_head_k =
       HEAD_DIM / num_elems_per_128b<CacheT>();
@@ -608,30 +515,28 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
   if (q_len <= 0) {
     return;
   }
-  T cache_k_scale_reg[IsDynamicC8 ? num_frags_z * 2 : num_frags_y * 4];
-  T cache_v_scale_reg[IsDynamicC8 ? num_frags_z * 4 : num_frags_y * 2];
-  if constexpr (!IsDynamicC8) {
-    if constexpr (is_scale_channel_wise) {
-      int scale_col_base = threadIdx.x % 4 * 2 + kv_head_idx * HEAD_DIM;
-      const T *cache_k_scale_cur_head = cache_k_scale + scale_col_base;
-      for (int i = 0; i < num_frags_y; ++i) {
-        const int scale_idx = i * 16;
-        cache_k_scale_reg[i * 4] = cache_k_scale_cur_head[scale_idx];
-        cache_k_scale_reg[i * 4 + 1] = cache_k_scale_cur_head[scale_idx + 1];
-        cache_k_scale_reg[i * 4 + 2] = cache_k_scale_cur_head[scale_idx + 8];
-        cache_k_scale_reg[i * 4 + 3] = cache_k_scale_cur_head[scale_idx + 9];
-      }
-      scale_col_base = threadIdx.x / 4 + kv_head_idx * HEAD_DIM;
-      const T *cache_v_scale_cur_head = cache_v_scale + scale_col_base;
-      for (int i = 0; i < num_frags_y; ++i) {
-        const int scale_idx = i * 16;
-        cache_v_scale_reg[i * 2] = cache_v_scale_cur_head[scale_idx];
-        cache_v_scale_reg[i * 2 + 1] = cache_v_scale_cur_head[scale_idx + 8];
-      }
-    } else {
-      cache_k_scale_reg[0] = cache_k_scale[kv_head_idx];
-      cache_v_scale_reg[0] = cache_v_scale[kv_head_idx];
+  T cache_k_scale_reg[num_frags_y * 4];
+  T cache_v_scale_reg[num_frags_y * 2];
+  if (is_scale_channel_wise) {
+    int scale_col_base = threadIdx.x % 4 * 2 + kv_head_idx * HEAD_DIM;
+    const T *cache_k_scale_cur_head = cache_k_scale + scale_col_base;
+    for (int i = 0; i < num_frags_y; ++i) {
+      const int scale_idx = i * 16;
+      cache_k_scale_reg[i * 4] = cache_k_scale_cur_head[scale_idx];
+      cache_k_scale_reg[i * 4 + 1] = cache_k_scale_cur_head[scale_idx + 1];
+      cache_k_scale_reg[i * 4 + 2] = cache_k_scale_cur_head[scale_idx + 8];
+      cache_k_scale_reg[i * 4 + 3] = cache_k_scale_cur_head[scale_idx + 9];
     }
+    scale_col_base = threadIdx.x / 4 + kv_head_idx * HEAD_DIM;
+    const T *cache_v_scale_cur_head = cache_v_scale + scale_col_base;
+    for (int i = 0; i < num_frags_y; ++i) {
+      const int scale_idx = i * 16;
+      cache_v_scale_reg[i * 2] = cache_v_scale_cur_head[scale_idx];
+      cache_v_scale_reg[i * 2 + 1] = cache_v_scale_cur_head[scale_idx + 8];
+    }
+  } else {
+    cache_k_scale_reg[0] = cache_k_scale[kv_head_idx];
+    cache_v_scale_reg[0] = cache_v_scale[kv_head_idx];
   }
   const uint32_t q_end =
       min(q_len, div_up((tile_id + 1) * num_rows_per_block, GROUP_SIZE));
@@ -670,7 +575,8 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
   const uint32_t kv_h_stride = BLOCK_SIZE * HEAD_DIM;
   const uint32_t kv_b_stride = HEAD_DIM;
   const uint32_t kv_d_stride = BLOCK_SIZE;
-  const uint32_t q_start_seq_id = cu_seqlens_q[batch_id];
+  const uint32_t q_start_seq_id =
+      batch_id * max_seq_len - __ldg(&cum_offsets[batch_id]);
   const uint32_t q_base_seq_id_this_block = tile_id * num_frags_x * 16;
   const uint32_t q_offset = q_start_seq_id * q_ori_n_stride +
                             q_head_idx * HEAD_DIM +
@@ -697,7 +603,7 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
           tid % 8 * num_elems_per_128b<T>();
     }
   }
-  const int *mask_offset_this_seq = mask_offset ? mask_offset + q_start_seq_id : nullptr;
+
   smem_t qo_smem(smem);
 
   uint32_t q_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head>(
@@ -722,17 +628,6 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
   smem_t k_smem(smem + num_frags_x * 16 * HEAD_DIM * sizeof(T)),
       v_smem(smem + num_frags_x * 16 * HEAD_DIM * sizeof(T) +
              NUM_WARP_KV * num_frags_z * 16 * HEAD_DIM * sizeof(CacheT));
-  T* k_smem_scale_ptr = nullptr;
-  T* v_smem_scale_ptr = nullptr;
-  smem_t k_scale_smem;
-  smem_t v_scale_smem;
-  if constexpr (IsDynamicC8) {
-    k_smem_scale_ptr = reinterpret_cast<T*>(smem + num_frags_x * 16 * HEAD_DIM * sizeof(T) +
-                                            NUM_WARP_KV * num_frags_z * 16 * HEAD_DIM * sizeof(CacheT) * 2);
-    v_smem_scale_ptr = k_smem_scale_ptr + NUM_WARP_KV * num_frags_z * 16;
-    k_scale_smem.base = reinterpret_cast<b128_t*>(k_smem_scale_ptr);
-    v_scale_smem.base = reinterpret_cast<b128_t*>(v_smem_scale_ptr);
-  }
 
   const uint32_t num_iterations = div_up(
       CAUSAL
@@ -749,7 +644,7 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
                          kv_len - q_len +
                              tile_id * num_rows_per_block / GROUP_SIZE,
                          chunk_start)))
-              : mask_offset ? 0 : chunk_len) /
+              : chunk_len) /
       (NUM_WARP_KV * num_frags_z * 16);
 
   uint32_t k_smem_offset_r =
@@ -795,20 +690,6 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
                                      kv_idx_base,
                                      chunk_end,
                                      const_k_offset);
-  if constexpr (IsDynamicC8) {
-    produce_kv_dynamic_scale_gmem2smem_async<SharedMemFillMode::kFillZero,
-                                             BLOCK_SIZE,
-                                             num_frags_z,
-                                             NUM_WARP_Q>(
-      k_scale_smem,
-      block_table_now,
-      cache_k_scale,
-      kv_idx_base,
-      kv_num_heads,
-      kv_head_idx,
-      chunk_end
-    );
-  }
   commit_group();
   produce_v_blockwise_c8<SharedMemFillMode::kNoFill,
                          NUM_WARPS,
@@ -826,34 +707,14 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
                                      kv_idx_base,
                                      chunk_end,
                                      const_v_offset);
-  if constexpr (IsDynamicC8) {
-    produce_kv_dynamic_scale_gmem2smem_async<SharedMemFillMode::kFillZero,
-                                             BLOCK_SIZE,
-                                             num_frags_z,
-                                             NUM_WARP_Q>(
-      v_scale_smem,
-      block_table_now,
-      cache_v_scale,
-      kv_idx_base,
-      kv_num_heads,
-      kv_head_idx,
-      chunk_end
-    );
-  }
   commit_group();
 #pragma unroll 1
   for (uint32_t iter = 0; iter < num_iterations; ++iter) {
     wait_group<1>();
     __syncthreads();
-    if constexpr (IsDynamicC8) {
-      produce_k_dynamic_scale_smem2reg<BLOCK_SIZE, num_frags_z, NUM_WARP_Q, T>(
-        k_smem_scale_ptr,
-        cache_k_scale_reg
-      );
-    }
 
     // s = qk
-    compute_qk_c8<num_frags_x, num_frags_y, num_frags_z, T, CacheT, is_scale_channel_wise, IsFP8, IsDynamicC8>(
+    compute_qk_c8<num_frags_x, num_frags_y, num_frags_z, T, CacheT, is_scale_channel_wise, IsFP8>(
         &qo_smem,
         &q_smem_offset_r,
         &k_smem,
@@ -869,16 +730,12 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
              NUM_WARPS,
              num_frags_x,
              num_frags_y,
-             num_frags_z>(attn_mask ? attn_mask + batch_id * attn_mask_len *attn_mask_len : nullptr,
-                          q_base_seq_id_this_block,
+             num_frags_z>(q_base_seq_id_this_block,
                           kv_idx_base + wid * num_frags_z * 16,
                           q_len,
                           kv_len,
                           chunk_end,
-                          attn_mask_len,
-                          s_frag,
-                          mask_offset_this_seq);
-
+                          s_frag);
     }
 
     // update m,d
@@ -886,7 +743,6 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
         s_frag, o_frag, m_frag, d_frag);
     __syncthreads();
 
-    const uint32_t ori_kv_idx_base = kv_idx_base;
     kv_idx_base += NUM_WARP_KV * num_frags_z * 16;
     produce_k_blockwise_c8<SharedMemFillMode::kNoFill,
                            NUM_WARPS,
@@ -904,29 +760,9 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
                                        kv_idx_base,
                                        chunk_end,
                                        const_k_offset);
-    if constexpr (IsDynamicC8) {
-      produce_kv_dynamic_scale_gmem2smem_async<SharedMemFillMode::kFillZero,
-                                              BLOCK_SIZE,
-                                              num_frags_z,
-                                              NUM_WARP_Q>(
-        k_scale_smem,
-        block_table_now,
-        cache_k_scale,
-        kv_idx_base,
-        kv_num_heads,
-        kv_head_idx,
-        chunk_end
-      );
-    }
     commit_group();
     wait_group<1>();
     __syncthreads();
-    if constexpr (IsDynamicC8) {
-      produce_v_dynamic_scale_smem2reg<BLOCK_SIZE, num_frags_z, NUM_WARP_Q, T>(
-        v_smem_scale_ptr,
-        cache_v_scale_reg
-      );
-    }
 
     // compute sfm * v
     compute_sfm_v_c8_iter_sq_bvec<num_frags_x,
@@ -935,9 +771,7 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
                                   BLOCK_SIZE,
                                   T,
                                   CacheT,
-                                  is_scale_channel_wise,
-                                  IsFP8,
-                                  IsDynamicC8>(
+                                  is_scale_channel_wise, IsFP8>(
         &v_smem, &v_smem_offset_r, s_frag, o_frag, d_frag, cache_v_scale_reg);
     __syncthreads();
 
@@ -957,20 +791,6 @@ __global__ void multi_query_append_attention_c8_warp1_4_kernel(
                                        kv_idx_base,
                                        chunk_end,
                                        const_v_offset);
-    if constexpr (IsDynamicC8) {
-      produce_kv_dynamic_scale_gmem2smem_async<SharedMemFillMode::kFillZero,
-                                              BLOCK_SIZE,
-                                              num_frags_z,
-                                              NUM_WARP_Q>(
-        v_scale_smem,
-        block_table_now,
-        cache_v_scale,
-        kv_idx_base,
-        kv_num_heads,
-        kv_head_idx,
-        chunk_end
-      );
-    }
     commit_group();
   }
   wait_group<0>();
@@ -1065,8 +885,7 @@ template <typename T,
           uint32_t NUM_WARP_Q,
           typename OutT = T,
           bool ENABLE_PREFILL = true,
-          bool IsFP8 = false,
-          bool IsDynamicC8 = false>
+          bool IsFP8=false>
 void MultiQueryAppendC8Attention(
     const AppendAttnMetaData &meta_data,
     const paddle::Tensor &qkv,
@@ -1080,8 +899,8 @@ void MultiQueryAppendC8Attention(
     const paddle::Tensor &seq_lens_q,
     const paddle::Tensor &seq_lens_kv,
     const paddle::Tensor &seq_lens_encoder,
-    const paddle::Tensor &batch_id_per_token,
-    const paddle::Tensor &cu_seqlens_q,
+    const paddle::Tensor &padding_offsets,
+    const paddle::Tensor &cum_offsets,
     const paddle::Tensor &block_table,
     const paddle::Tensor &batch_ids,
     const paddle::Tensor &tile_ids_per_batch,
@@ -1124,8 +943,7 @@ void MultiQueryAppendC8Attention(
     constexpr uint32_t num_frags_z = BLOCK_SIZE / 16;
     constexpr uint32_t smem_size =
         num_warps * num_frags_x * 16 * HEAD_DIM * sizeof(T) +
-        num_frags_z * 16 * HEAD_DIM * sizeof(uint8_t) * 2 +
-        num_frags_z * 16 * sizeof(T) * 2;
+        num_frags_z * 16 * HEAD_DIM * sizeof(uint8_t) * 2;
     auto split_kv_kernel =
         multi_query_append_attention_c8_kernel<NV_TYPE,
                                                uint8_t,
@@ -1142,9 +960,7 @@ void MultiQueryAppendC8Attention(
                                                num_frags_y,
                                                OUT_NV_TYPE,
                                                ENABLE_PREFILL,
-                                               false,
-                                               IsFP8,
-                                               IsDynamicC8>;
+                                               false, IsFP8>;
     if (is_scale_channel_wise) {
       split_kv_kernel =
         multi_query_append_attention_c8_kernel<NV_TYPE,
@@ -1162,9 +978,7 @@ void MultiQueryAppendC8Attention(
                                                num_frags_y,
                                                OUT_NV_TYPE,
                                                ENABLE_PREFILL,
-                                               true,
-                                               IsFP8,
-                                               IsDynamicC8>;
+                                               true, IsFP8>;
     }
     if (smem_size >= 48 * 1024) {
       cudaFuncSetAttribute(split_kv_kernel,
@@ -1198,9 +1012,7 @@ void MultiQueryAppendC8Attention(
                                                  num_frags_y,
                                                  OUT_NV_TYPE,
                                                  ENABLE_PREFILL,
-                                                 false,
-                                                 IsFP8,
-                                                 IsDynamicC8>;
+                                                 false, IsFP8>;
       if (is_scale_channel_wise) {
         nosplit_kv_kernel =
           multi_query_append_attention_c8_kernel<NV_TYPE,
@@ -1218,9 +1030,7 @@ void MultiQueryAppendC8Attention(
                                                  num_frags_y,
                                                  OUT_NV_TYPE,
                                                  ENABLE_PREFILL,
-                                                 true,
-                                                 IsFP8,
-                                                 IsDynamicC8>;
+                                                 true, IsFP8>;
       }
       if (smem_size >= 48 * 1024) {
         cudaFuncSetAttribute(nosplit_kv_kernel,
@@ -1244,9 +1054,8 @@ void MultiQueryAppendC8Attention(
           seq_lens_kv.data<int>(),
           batch_ids.data<int>(),
           tile_ids_per_batch.data<int>(),
-          cu_seqlens_q.data<int>(),
+          cum_offsets.data<int>(),
           block_table.data<int>(),
-          meta_data.mask_offset,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1302,9 +1111,8 @@ void MultiQueryAppendC8Attention(
           seq_lens_kv.data<int>(),
           batch_ids.data<int>(),
           tile_ids_per_batch.data<int>(),
-          cu_seqlens_q.data<int>(),
+          cum_offsets.data<int>(),
           block_table.data<int>(),
-          meta_data.mask_offset,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1338,7 +1146,7 @@ void MultiQueryAppendC8Attention(
                 seq_lens_q.data<int>(),
                 seq_lens_kv.data<int>(),
                 seq_lens_encoder.data<int>(),
-                cu_seqlens_q.data<int>(),
+                cum_offsets.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
                                  const_cast<T *>(shift_bias.get().data<T>()))
                            : nullptr,
@@ -1373,8 +1181,7 @@ void MultiQueryAppendC8Attention(
                 seq_lens_q.data<int>(),
                 seq_lens_kv.data<int>(),
                 seq_lens_encoder.data<int>(),
-                batch_id_per_token.data<int>(),
-                cu_seqlens_q.data<int>(),
+                padding_offsets.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
                                  const_cast<T *>(shift_bias.get().data<T>()))
                            : nullptr,
@@ -1398,8 +1205,7 @@ void MultiQueryAppendC8Attention(
     constexpr uint32_t num_frags_z = BLOCK_SIZE / 16 / NUM_WARP_KV * 2;
     constexpr uint32_t smem_size =
         num_frags_x * 16 * HEAD_DIM * sizeof(T) +
-        NUM_WARP_KV * num_frags_z * 16 * HEAD_DIM * sizeof(uint8_t) * 2 +
-        NUM_WARP_KV * num_frags_z * 16 * sizeof(T) * 2;
+        NUM_WARP_KV * num_frags_z * 16 * HEAD_DIM * sizeof(uint8_t) * 2;
     auto split_kv_kernel =
         multi_query_append_attention_c8_warp1_4_kernel<NV_TYPE,
                                                        uint8_t,
@@ -1416,9 +1222,7 @@ void MultiQueryAppendC8Attention(
                                                        num_frags_y,
                                                        OUT_NV_TYPE,
                                                        ENABLE_PREFILL,
-                                                       false,
-                                                       IsFP8,
-                                                       IsDynamicC8>;
+                                                       false, IsFP8>;
     if (is_scale_channel_wise) {
       split_kv_kernel =
         multi_query_append_attention_c8_warp1_4_kernel<NV_TYPE,
@@ -1436,9 +1240,7 @@ void MultiQueryAppendC8Attention(
                                                        num_frags_y,
                                                        OUT_NV_TYPE,
                                                        ENABLE_PREFILL,
-                                                       true,
-                                                       IsFP8,
-                                                       IsDynamicC8>;
+                                                       true, IsFP8>;
     }
     if (smem_size >= 48 * 1024) {
       cudaFuncSetAttribute(split_kv_kernel,
@@ -1453,17 +1255,10 @@ void MultiQueryAppendC8Attention(
       chunk_size = static_cast<uint32_t>(encoder_max_partition_size);
     }
 
-    const int num_chunks = div_up(max_seq_len, chunk_size);
-    uint32_t attn_mask_len;
-    if (attn_mask) {
-        attn_mask_len = attn_mask.get().shape()[1];
-    } else {
-        attn_mask_len = -1;
-    }
-
+    const int num_chunks = div_up(max_dec_len, chunk_size);
     dim3 grids(num_blocks_x_cpu, num_chunks, kv_num_heads);
     dim3 blocks(32, num_warps);
-    if (num_chunks <= 0) {
+    if (num_chunks <= 1) {
       auto nosplit_kv_kernel =
           multi_query_append_attention_c8_warp1_4_kernel<NV_TYPE,
                                                          uint8_t,
@@ -1480,9 +1275,7 @@ void MultiQueryAppendC8Attention(
                                                          num_frags_y,
                                                          OUT_NV_TYPE,
                                                          ENABLE_PREFILL,
-                                                         false,
-                                                         IsFP8,
-                                                         IsDynamicC8>;
+                                                         false, IsFP8>;
       if (is_scale_channel_wise) {
         nosplit_kv_kernel =
           multi_query_append_attention_c8_warp1_4_kernel<NV_TYPE,
@@ -1500,9 +1293,7 @@ void MultiQueryAppendC8Attention(
                                                          num_frags_y,
                                                          OUT_NV_TYPE,
                                                          ENABLE_PREFILL,
-                                                         true,
-                                                         IsFP8,
-                                                         IsDynamicC8>;
+                                                         true, IsFP8>;
       }
       if (smem_size >= 48 * 1024) {
         cudaFuncSetAttribute(nosplit_kv_kernel,
@@ -1526,11 +1317,8 @@ void MultiQueryAppendC8Attention(
           seq_lens_kv.data<int>(),
           batch_ids.data<int>(),
           tile_ids_per_batch.data<int>(),
-          cu_seqlens_q.data<int>(),
+          cum_offsets.data<int>(),
           block_table.data<int>(),
-          meta_data.mask_offset,
-          attn_mask ? const_cast<bool *>(attn_mask.get().data<bool>())
-                        : nullptr,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1543,8 +1331,7 @@ void MultiQueryAppendC8Attention(
           nullptr,
           nullptr,
           reinterpret_cast<OUT_NV_TYPE *>(out->data<OutT>()),
-          speculate_max_draft_token_num,
-          attn_mask_len);
+          speculate_max_draft_token_num);
     } else {
       phi::Allocator::AllocationPtr tmp_workspace, tmp_m, tmp_d;
       if (is_decoder) {
@@ -1591,8 +1378,8 @@ void MultiQueryAppendC8Attention(
           reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_k_scale.data<T>())),
           reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_v_scale.data<T>())),
           shift_bias ? reinterpret_cast<NV_TYPE *>(
-                            const_cast<T *>(shift_bias.get().data<T>()))
-                      : nullptr,
+                           const_cast<T *>(shift_bias.get().data<T>()))
+                     : nullptr,
           smooth_weight ? reinterpret_cast<NV_TYPE *>(
                               const_cast<T *>(smooth_weight.get().data<T>()))
                         : nullptr,
@@ -1600,11 +1387,8 @@ void MultiQueryAppendC8Attention(
           seq_lens_kv.data<int>(),
           batch_ids.data<int>(),
           tile_ids_per_batch.data<int>(),
-          cu_seqlens_q.data<int>(),
+          cum_offsets.data<int>(),
           block_table.data<int>(),
-          meta_data.mask_offset,
-          attn_mask ? const_cast<bool *>(attn_mask.get().data<bool>())
-                        : nullptr,
           max_seq_len,
           max_dec_len,
           max_block_num_per_seq,
@@ -1617,8 +1401,7 @@ void MultiQueryAppendC8Attention(
           static_cast<float *>(tmp_m->ptr()),
           static_cast<float *>(tmp_d->ptr()),
           reinterpret_cast<OUT_NV_TYPE *>(out->data<OutT>()),
-          speculate_max_draft_token_num,
-          attn_mask_len);
+          speculate_max_draft_token_num);
       // merge
       constexpr int vec_size = num_elems_per_128b<NV_TYPE>();
       if (is_decoder) {
@@ -1634,10 +1417,10 @@ void MultiQueryAppendC8Attention(
                 seq_lens_q.data<int>(),
                 seq_lens_kv.data<int>(),
                 seq_lens_encoder.data<int>(),
-                cu_seqlens_q.data<int>(),
+                cum_offsets.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
-                                  const_cast<T *>(shift_bias.get().data<T>()))
-                            : nullptr,
+                                 const_cast<T *>(shift_bias.get().data<T>()))
+                           : nullptr,
                 smooth_weight ? reinterpret_cast<NV_TYPE *>(const_cast<T *>(
                                     smooth_weight.get().data<T>()))
                               : nullptr,
@@ -1654,14 +1437,14 @@ void MultiQueryAppendC8Attention(
         constexpr int blockx = HEAD_DIM / vec_size;
         constexpr int blocky = (128 + blockx - 1) / blockx;
         dim3 grids_merge(min(sm_count * 4, token_num),
-                          num_heads);
+                         num_heads);
         dim3 blocks_merge(blockx, blocky);
         merge_multi_chunks_v2_kernel<NV_TYPE,
-                                      vec_size,
-                                      blocky,
-                                      HEAD_DIM,
-                                      OUT_NV_TYPE,
-                                      ENABLE_PREFILL>
+                                     vec_size,
+                                     blocky,
+                                     HEAD_DIM,
+                                     OUT_NV_TYPE,
+                                     ENABLE_PREFILL>
             <<<grids_merge, blocks_merge, 0, stream>>>(
                 reinterpret_cast<NV_TYPE *>(tmp_workspace->ptr()),
                 static_cast<float *>(tmp_m->ptr()),
@@ -1669,11 +1452,10 @@ void MultiQueryAppendC8Attention(
                 seq_lens_q.data<int>(),
                 seq_lens_kv.data<int>(),
                 seq_lens_encoder.data<int>(),
-                batch_id_per_token.data<int>(),
-                cu_seqlens_q.data<int>(),
+                padding_offsets.data<int>(),
                 shift_bias ? reinterpret_cast<NV_TYPE *>(
-                                  const_cast<T *>(shift_bias.get().data<T>()))
-                            : nullptr,
+                                 const_cast<T *>(shift_bias.get().data<T>()))
+                           : nullptr,
                 smooth_weight ? reinterpret_cast<NV_TYPE *>(const_cast<T *>(
                                     smooth_weight.get().data<T>()))
                               : nullptr,
@@ -1717,8 +1499,8 @@ void CascadeAppendAttentionC8Kernel(
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,
@@ -1735,7 +1517,6 @@ void CascadeAppendAttentionC8Kernel(
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
-    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out) {
   const auto token_num = meta_data.token_nums;
@@ -1744,7 +1525,6 @@ void CascadeAppendAttentionC8Kernel(
   const auto num_heads = meta_data.q_num_heads;
   const auto group_size = meta_data.q_num_heads / meta_data.kv_num_heads;
   const auto head_dim = meta_data.head_dims;
-  bool is_dynamic_cfp8 = cache_quant_type_str == "block_wise_fp8";
 
   DISPATCH_CAUSAL(
       causal,
@@ -1763,46 +1543,43 @@ void CascadeAppendAttentionC8Kernel(
                       BLOCK_SIZE,
                       {DISPATCH_BLOCKSHAPE_Q(
                           block_shape_q, BLOCK_SHAPE_Q, NUM_WARP_Q, {
-                            DISPATCH_DyCfp8(is_dynamic_cfp8, IsDynamicC8, {
-                              MultiQueryAppendC8Attention<T,
-                                                          GROUP_SIZE,
-                                                          HEAD_DIM,
-                                                          BLOCK_SIZE,
-                                                          CAUSAL,
-                                                          BLOCK_SHAPE_Q,
-                                                          NUM_WARP_Q,
-                                                          OutT,
-                                                          ENABLE_PREFILL,
-                                                          IsFP8,
-                                                          IsDynamicC8>(
-                                  meta_data,
-                                  qkv,
-                                  cache_k,
-                                  cache_v,
-                                  attn_mask,
-                                  cache_k_scale.get(),
-                                  cache_v_scale.get(),
-                                  shift_bias,
-                                  smooth_weight,
-                                  seq_lens_q,
-                                  seq_lens_kv,
-                                  seq_lens_encoder,
-                                  batch_id_per_token,
-                                  cu_seqlens_q,
-                                  block_table,
-                                  batch_ids,
-                                  tile_ids_per_batch,
-                                  num_blocks,
-                                  max_seq_len,
-                                  max_dec_len,
-                                  quant_max_bound,
-                                  quant_min_bound,
-                                  in_scale,
-                                  max_partition_size,
-                                  encoder_max_partition_size,
-                                  speculate_max_draft_token_num,
-                                  is_decoder,
-                                  stream,
-                                  out);
-      })})})})})})})
+                            MultiQueryAppendC8Attention<T,
+                                                        GROUP_SIZE,
+                                                        HEAD_DIM,
+                                                        BLOCK_SIZE,
+                                                        CAUSAL,
+                                                        BLOCK_SHAPE_Q,
+                                                        NUM_WARP_Q,
+                                                        OutT,
+                                                        ENABLE_PREFILL, IsFP8>(
+                                meta_data,
+                                qkv,
+                                cache_k,
+                                cache_v,
+                                attn_mask,
+                                cache_k_scale.get(),
+                                cache_v_scale.get(),
+                                shift_bias,
+                                smooth_weight,
+                                seq_lens_q,
+                                seq_lens_kv,
+                                seq_lens_encoder,
+                                padding_offsets,
+                                cum_offsets,
+                                block_table,
+                                batch_ids,
+                                tile_ids_per_batch,
+                                num_blocks,
+                                max_seq_len,
+                                max_dec_len,
+                                quant_max_bound,
+                                quant_min_bound,
+                                in_scale,
+                                max_partition_size,
+                                encoder_max_partition_size,
+                                speculate_max_draft_token_num,
+                                is_decoder,
+                                stream,
+                                out);
+                          })})})})})})
 }

custom_ops/gpu_ops/append_attn/append_attention_func.cuh

@@ -384,105 +384,6 @@ __device__ __forceinline__ void produce_v_blockwise_c8(
   }
 }
 
-template<SharedMemFillMode fill_mode,
-         uint32_t block_size,
-         uint32_t num_frags_z,
-         uint32_t NUM_WARP_Q,
-         typename T>
-__device__ __forceinline__ void produce_kv_dynamic_scale_gmem2smem_async(
-  smem_t kv_scale_smem,
-  const int* block_table_now,
-  const T* cache_kv_scale,
-  const uint32_t kv_idx,
-  const uint32_t kv_num_heads,
-  const uint32_t kv_head_idx,
-  const uint32_t chunk_end
-) {
-  const uint32_t tx = threadIdx.x, ty = threadIdx.y;
-  const uint32_t tid = ty * 32 + tx;
-  if constexpr (NUM_WARP_Q == 4) {
-    // 4 warps shared block_size
-    int block_id = __ldg(&block_table_now[kv_idx / block_size]);
-    if (block_id < 0) block_id = 0;
-    if (tid < block_size / 8) {
-      const T* cache_k_scale_now = cache_kv_scale + block_id * kv_num_heads * block_size + kv_head_idx * block_size + tid * 8;
-      const int kv_idx_this_thread = kv_idx + tid * 8;
-      kv_scale_smem.load_128b_async<fill_mode>(tid, cache_k_scale_now, kv_idx_this_thread < chunk_end);
-    }
-  } else {
-    // 1 warp 32 tokens
-    if (tid < block_size / 8 * 2) {
-      const uint32_t kv_idx_now = kv_idx + block_size * tid / 8;
-      int block_id = __ldg(&block_table_now[kv_idx_now / block_size]);
-      if (block_id < 0) block_id = 0;
-      const int kv_idx_this_thread = kv_idx + tid * 8;
-      const T* cache_k_scale_now = cache_kv_scale + block_id * kv_num_heads * block_size + kv_head_idx * block_size + tid % 8 * 8;
-      kv_scale_smem.load_128b_async<fill_mode>(tid, cache_k_scale_now, kv_idx_this_thread < chunk_end);
-    }
-  }
-}
-
-template<uint32_t block_size,
-         uint32_t num_frags_z,
-         uint32_t NUM_WARP_Q,
-         typename T>
-__device__ __forceinline__ void produce_k_dynamic_scale_smem2reg(
-  T* k_smem_scale,
-  T* cache_k_reg
-) {
-  const uint32_t tx = threadIdx.x, ty = threadIdx.y;
-  if constexpr (NUM_WARP_Q == 4) {
-    // 4 warps shared block_size
-    const uint32_t row_id = tx / 4;
-    for (uint32_t fz = 0; fz < num_frags_z; fz++) {
-      const uint32_t scale_idx = fz * 16 + row_id;
-      cache_k_reg[fz * 2] = k_smem_scale[scale_idx];
-      cache_k_reg[fz * 2 + 1] = k_smem_scale[scale_idx + 8];
-    }
-  } else {
-    // 1 warp 32 tokens
-    const uint32_t row_id = tx / 4;
-    for (uint32_t fz = 0; fz < num_frags_z; fz++) {
-      const uint32_t scale_idx = ty * 32 + fz * 16 + row_id;
-      cache_k_reg[fz * 2] = k_smem_scale[scale_idx];
-      cache_k_reg[fz * 2 + 1] = k_smem_scale[scale_idx + 8];
-    }
-  }
-}
-
-template<uint32_t block_size,
-         uint32_t num_frags_z,
-         uint32_t NUM_WARP_Q,
-         typename T>
-__device__ __forceinline__ void produce_v_dynamic_scale_smem2reg(
-  T* v_smem_scale,
-  T* cache_v_reg
-) {
-  const uint32_t tx = threadIdx.x, ty = threadIdx.y;
-
-  if constexpr (NUM_WARP_Q == 4) {
-    // 4 warps shared block_size
-    const uint32_t row_id = tx % 4 * 2;
-    for (uint32_t fz = 0; fz < num_frags_z; fz++) {
-      const uint32_t scale_idx = fz * 16 + row_id;
-      cache_v_reg[fz * 4] = v_smem_scale[scale_idx];
-      cache_v_reg[fz * 4 + 1] = v_smem_scale[scale_idx + 1];
-      cache_v_reg[fz * 4 + 2] = v_smem_scale[scale_idx + 8];
-      cache_v_reg[fz * 4 + 3] = v_smem_scale[scale_idx + 9];
-    }
-  } else {
-    // 1 warp 32 tokens
-    const uint32_t row_id = tx % 4 * 2;
-    for (uint32_t fz = 0; fz < num_frags_z; fz++) {
-      const uint32_t scale_idx = ty * 32 + fz * 16 + row_id;
-      cache_v_reg[fz * 4] = v_smem_scale[scale_idx];
-      cache_v_reg[fz * 4 + 1] = v_smem_scale[scale_idx + 1];
-      cache_v_reg[fz * 4 + 2] = v_smem_scale[scale_idx + 8];
-      cache_v_reg[fz * 4 + 3] = v_smem_scale[scale_idx + 9];
-    }
-  }
-}
-
 template <SharedMemFillMode fill_mode,
           uint32_t num_warps,
           uint32_t block_size,
@@ -915,8 +816,7 @@ template <uint32_t num_frags_x,
           typename T,
           typename CacheT,
           bool is_scale_channel_wise = false,
-          bool IsFP8 = false,
-          bool IsDynamicC8 = false>
+          bool IsFP8=false>
 __device__ __forceinline__ void compute_qk_c8(smem_t* q_smem,
                                               uint32_t* q_smem_offset_r,
                                               smem_t* k_smem,
@@ -960,27 +860,20 @@ __device__ __forceinline__ void compute_qk_c8(smem_t* q_smem,
         convert_c8<T,IsFP8>(b_frag_dq_T, b_frag[fy * 2]);
         convert_c8<T,IsFP8>(b_frag_dq_T + 4, b_frag[fy * 2 + 1]);
         // scale zp
-        if constexpr (!IsDynamicC8) {
-          if constexpr (is_scale_channel_wise) {
-            const int scale_col = (ky * 2 + fy) * 4;
-            b_frag_dq_T[0] *= cache_k_scale[scale_col];
-            b_frag_dq_T[1] *= cache_k_scale[scale_col + 1];
-            b_frag_dq_T[2] *= cache_k_scale[scale_col + 2];
-            b_frag_dq_T[3] *= cache_k_scale[scale_col + 3];
-            b_frag_dq_T[4] *= cache_k_scale[scale_col];
-            b_frag_dq_T[5] *= cache_k_scale[scale_col + 1];
-            b_frag_dq_T[6] *= cache_k_scale[scale_col + 2];
-            b_frag_dq_T[7] *= cache_k_scale[scale_col + 3];
-          } else {
-#pragma unroll
-            for (uint32_t b_i = 0; b_i < 8; ++b_i) {
-              b_frag_dq_T[b_i] *= cache_k_scale[0];
-            }
-          }
+        if constexpr (is_scale_channel_wise) {
+          const int scale_col = (ky * 2 + fy) * 4;
+          b_frag_dq_T[0] *= cache_k_scale[scale_col];
+          b_frag_dq_T[1] *= cache_k_scale[scale_col + 1];
+          b_frag_dq_T[2] *= cache_k_scale[scale_col + 2];
+          b_frag_dq_T[3] *= cache_k_scale[scale_col + 3];
+          b_frag_dq_T[4] *= cache_k_scale[scale_col];
+          b_frag_dq_T[5] *= cache_k_scale[scale_col + 1];
+          b_frag_dq_T[6] *= cache_k_scale[scale_col + 2];
+          b_frag_dq_T[7] *= cache_k_scale[scale_col + 3];
         } else {
 #pragma unroll
           for (uint32_t b_i = 0; b_i < 8; ++b_i) {
-            b_frag_dq_T[b_i] *= cache_k_scale[fz * 2 + b_i / 4];
+            b_frag_dq_T[b_i] *= cache_k_scale[0];
           }
         }
 #pragma unroll
@@ -1012,15 +905,12 @@ template <typename T,
           uint32_t num_frags_y,
           uint32_t num_frags_z,
           bool IS_SYSTEM = false>
-__device__ __forceinline__ void mask_s(const bool* attn_mask,
-                                       const uint32_t qo_idx_base,
+__device__ __forceinline__ void mask_s(const uint32_t qo_idx_base,
                                        const uint32_t kv_idx_base,
                                        const uint32_t qo_len,
                                        const uint32_t kv_len,
                                        const uint32_t chunk_end,
-                                       const uint32_t attn_mask_len,
-                                       float (*s_frag)[num_frags_z][8],
-                                       const int *mask_offset = nullptr) {
+                                       float (*s_frag)[num_frags_z][8]) {
   const uint32_t tx = threadIdx.x;
 #pragma unroll
   for (uint32_t fx = 0; fx < num_frags_x; ++fx) {
@@ -1034,21 +924,10 @@ __device__ __forceinline__ void mask_s(const bool* attn_mask,
                                  group_size,
                          kv_idx = kv_idx_base + fz * 16 + 2 * (tx % 4) +
                                   8 * (reg_id / 4) + reg_id % 2;
-          bool out_of_boundary;
-          if (mask_offset) {
-            out_of_boundary = q_idx < qo_len ? (kv_idx > mask_offset[q_idx]) : true;
-          } else {
-            out_of_boundary =
-                (causal
-                    ? (kv_idx > kv_len + q_idx - qo_len || (kv_idx >= chunk_end))
-                    : kv_idx >= chunk_end);
-            if (attn_mask != nullptr && kv_idx > kv_len - qo_len && kv_idx < chunk_end && q_idx < attn_mask_len) {
-              const int32_t mask_idx = q_idx * attn_mask_len + kv_idx - kv_len + qo_len;
-              bool mask = attn_mask[mask_idx];
-              out_of_boundary |= mask;
-            }
-          }
-
+          const bool out_of_boundary =
+              (causal
+                   ? (kv_idx > kv_len + q_idx - qo_len || (kv_idx >= chunk_end))
+                   : kv_idx >= chunk_end);
           if constexpr (std::is_same<T, half>::value) {
             s_frag[fx][fz][reg_id] =
                 out_of_boundary ? -5e4f : s_frag[fx][fz][reg_id];
@@ -1056,7 +935,6 @@ __device__ __forceinline__ void mask_s(const bool* attn_mask,
             s_frag[fx][fz][reg_id] =
                 out_of_boundary ? -3.0e+30f : s_frag[fx][fz][reg_id];
           }
-          // printf("tid: %d. qk[%u,%u] = %f, mask: %d \n ", threadIdx.x, kv_idx, q_idx, static_cast<float>(s_frag[fx][fz][reg_id]), int(out_of_boundary));
         } else {
           const uint32_t q_idx = qo_idx_base,
                          kv_idx = kv_idx_base + fz * 16 + 2 * (tx % 4) +
@@ -1200,9 +1078,7 @@ template <uint32_t num_frags_x,
           uint32_t block_size,
           typename T,
           typename CacheT,
-          bool is_scale_channel_wise = false,
-          bool IsFP8 = false,
-          bool IsDynamicC8 = false>
+          bool is_scale_channel_wise = false, bool IsFP8=false>
 __device__ __forceinline__ void compute_sfm_v_c8(
     smem_t* v_smem,
     uint32_t* v_smem_offset_r,
@@ -1244,28 +1120,16 @@ __device__ __forceinline__ void compute_sfm_v_c8(
         convert_c8<T,IsFP8>(b_frag_dq_T, b_frag[fz * 2]);
         convert_c8<T,IsFP8>(b_frag_dq_T + 4, b_frag[fz * 2 + 1]);
         // scale zp
-        if constexpr (!IsDynamicC8) {
-          if constexpr (is_scale_channel_wise) {
+        if constexpr (is_scale_channel_wise) {
 #pragma unroll
-            for (uint32_t b_i = 0; b_i < 8; ++b_i) {
-              b_frag_dq_T[b_i] *= cache_v_scale[b_i / 4 + fy * 2];
-            }
-          } else {
-#pragma unroll
-            for (uint32_t b_i = 0; b_i < 8; ++b_i) {
-              b_frag_dq_T[b_i] *= cache_v_scale[0];
-            }
+          for (uint32_t b_i = 0; b_i < 8; ++b_i) {
+            b_frag_dq_T[b_i] *= cache_v_scale[b_i / 4 + fy * 2];
           }
         } else {
-          const int scale_col = (kz * 2 + fz) * 4;
-          b_frag_dq_T[0] *= cache_v_scale[scale_col];
-          b_frag_dq_T[1] *= cache_v_scale[scale_col + 1];
-          b_frag_dq_T[2] *= cache_v_scale[scale_col + 2];
-          b_frag_dq_T[3] *= cache_v_scale[scale_col + 3];
-          b_frag_dq_T[4] *= cache_v_scale[scale_col];
-          b_frag_dq_T[5] *= cache_v_scale[scale_col + 1];
-          b_frag_dq_T[6] *= cache_v_scale[scale_col + 2];
-          b_frag_dq_T[7] *= cache_v_scale[scale_col + 3];
+#pragma unroll
+          for (uint32_t b_i = 0; b_i < 8; ++b_i) {
+            b_frag_dq_T[b_i] *= cache_v_scale[0];
+          }
         }
 #pragma unroll
         for (uint32_t fx = 0; fx < num_frags_x; ++fx) {  // m: num_frags_x * 16
@@ -1292,9 +1156,7 @@ template <uint32_t num_frags_x,
           uint32_t block_size,
           typename T,
           typename CacheT,
-          bool is_scale_channel_wise = false,
-          bool IsFP8 = false,
-          bool IsDynamicC8 = false>
+          bool is_scale_channel_wise = false, bool IsFP8=false>
 __device__ __forceinline__ void compute_sfm_v_c8_iter_sq_bvec(
     smem_t* v_smem,
     uint32_t* v_smem_offset_r,
@@ -1338,28 +1200,16 @@ __device__ __forceinline__ void compute_sfm_v_c8_iter_sq_bvec(
         convert_c8<T,IsFP8>(b_frag_dq_T, b_frag[fz * 2]);
         convert_c8<T,IsFP8>(b_frag_dq_T + 4, b_frag[fz * 2 + 1]);
         // scale zp
-        if constexpr (!IsDynamicC8) {
-          if constexpr (is_scale_channel_wise) {
+        if constexpr (is_scale_channel_wise) {
 #pragma unroll
-            for (uint32_t b_i = 0; b_i < 8; ++b_i) {
-              b_frag_dq_T[b_i] *= cache_v_scale[b_i / 4 + fy * 2];
-            }
-          } else {
-            #pragma unroll
-            for (uint32_t b_i = 0; b_i < 8; ++b_i) {
-              b_frag_dq_T[b_i] *= cache_v_scale[0];
-            }
+          for (uint32_t b_i = 0; b_i < 8; ++b_i) {
+            b_frag_dq_T[b_i] *= cache_v_scale[b_i / 4 + fy * 2];
           }
         } else {
-          const int scale_col = (kz * 2 + fz) * 4;
-          b_frag_dq_T[0] *= cache_v_scale[scale_col];
-          b_frag_dq_T[1] *= cache_v_scale[scale_col + 1];
-          b_frag_dq_T[2] *= cache_v_scale[scale_col + 2];
-          b_frag_dq_T[3] *= cache_v_scale[scale_col + 3];
-          b_frag_dq_T[4] *= cache_v_scale[scale_col];
-          b_frag_dq_T[5] *= cache_v_scale[scale_col + 1];
-          b_frag_dq_T[6] *= cache_v_scale[scale_col + 2];
-          b_frag_dq_T[7] *= cache_v_scale[scale_col + 3];
+          #pragma unroll
+          for (uint32_t b_i = 0; b_i < 8; ++b_i) {
+            b_frag_dq_T[b_i] *= cache_v_scale[0];
+          }
         }
 #pragma unroll
         for (uint32_t fx = 0; fx < num_frags_x; ++fx) {  // m: num_frags_x * 16
@@ -2002,7 +1852,7 @@ __global__ void merge_multi_chunks_kernel(
     const float* __restrict__ multi_d,  // [token_num, num_chunks, num_heads]
     const int* __restrict__ seq_lens_q,
     const int* __restrict__ seq_lens_kv,
-    const int* __restrict__ batch_id_per_token,
+    const int* __restrict__ padding_offsets,
     const T* __restrict__ shift_bias,     // [q_num_heads * HEAD_DIM]
     const T* __restrict__ smooth_weight,  // [q_num_heads * HEAD_DIM]
     T* __restrict__ out,
@@ -2016,7 +1866,8 @@ __global__ void merge_multi_chunks_kernel(
     const int head_dim) {
   const int vid = threadIdx.x, hid = threadIdx.y;
   const int qid = blockIdx.x;
-  const uint32_t bid = batch_id_per_token[qid];
+  const uint32_t ori_token_id = qid + padding_offsets[qid];
+  const uint32_t bid = ori_token_id / max_seq_len;
   if (seq_lens_q[bid] <= 0 || seq_lens_kv[bid] <= 0) {
     return;
   }
@@ -2260,7 +2111,7 @@ __global__ void merge_multi_chunks_decoder_kernel(
     const int *__restrict__ seq_lens_q,
     const int *__restrict__ seq_lens_kv,
     const int *__restrict__ seq_lens_encoder,
-    const int *__restrict__ cu_seqlens_q,
+    const int *__restrict__ cum_offsets,
     const T *__restrict__ shift_bias,     // [q_num_heads * HEAD_DIM]
     const T *__restrict__ smooth_weight,  // [q_num_heads * HEAD_DIM]
     OutT *__restrict__ out,
@@ -2276,7 +2127,7 @@ __global__ void merge_multi_chunks_decoder_kernel(
   const int bid = blockIdx.x, hid = blockIdx.y;
   __shared__ T smem[bdy * HEAD_DIM];
   __shared__ float md_smem[bdy * 2];
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - cum_offsets[bid];
   const int seq_len_q = seq_lens_q[bid];
   if (seq_len_q == 0) return;
   int seq_len_kv = seq_lens_kv[bid];
@@ -2389,8 +2240,7 @@ __global__ void merge_multi_chunks_v2_kernel(
     const int *__restrict__ seq_lens_q,
     const int *__restrict__ seq_lens_kv,
     const int *__restrict__ seq_lens_encoder,
-    const int *__restrict__ batch_id_per_token,
-    const int *__restrict__ cu_seqlens_q,
+    const int *__restrict__ padding_offsets,
     const T *__restrict__ shift_bias,     // [q_num_heads * HEAD_DIM]
     const T *__restrict__ smooth_weight,  // [q_num_heads * HEAD_DIM]
     OutT *__restrict__ out,
@@ -2409,8 +2259,9 @@ __global__ void merge_multi_chunks_v2_kernel(
   __shared__ T smem[bdy * HEAD_DIM];
   __shared__ float md_smem[bdy * 2];
   for (int qid = blockIdx.x; qid < token_num; qid += gridDim.x) {
-    const uint32_t bid = batch_id_per_token[qid];
-    const uint32_t local_seq_id = qid - cu_seqlens_q[bid];
+    const uint32_t ori_token_id = qid + padding_offsets[qid];
+    const uint32_t bid = ori_token_id / max_seq_len;
+    const uint32_t local_seq_id = ori_token_id % max_seq_len;
     const int seq_len_q = seq_lens_q[bid];
     if (seq_len_q == 0) continue;
     int seq_len_kv = seq_lens_kv[bid];

custom_ops/gpu_ops/append_attn/append_attention_kernel.h

@@ -40,8 +40,8 @@ void CascadeAppendAttentionC16Kernel(
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,
@@ -85,8 +85,8 @@ void CascadeAppendAttentionC8Kernel(
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,
@@ -103,7 +103,6 @@ void CascadeAppendAttentionC8Kernel(
     const bool causal,
     const bool is_decoder,
     const bool enable_prefill,
-    const std::string& cache_quant_type_str,
     cudaStream_t& stream,
     paddle::Tensor* out);
 
@@ -131,8 +130,8 @@ void CascadeAppendAttentionC4Kernel(
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,
@@ -176,8 +175,8 @@ void CascadeAppendAttentionKernel(
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,
@@ -212,8 +211,8 @@ void CascadeAppendAttentionKernel(
                                                 seq_lens_q,
                                                 seq_lens_kv,
                                                 seq_lens_encoder,
-                                                batch_id_per_token,
-                                                cu_seqlens_q,
+                                                padding_offsets,
+                                                cum_offsets,
                                                 block_table,
                                                 batch_ids,
                                                 tile_ids_per_batch,
@@ -247,8 +246,8 @@ void CascadeAppendAttentionKernel(
                                                 seq_lens_q,
                                                 seq_lens_kv,
                                                 seq_lens_encoder,
-                                                batch_id_per_token,
-                                                cu_seqlens_q,
+                                                padding_offsets,
+                                                cum_offsets,
                                                 block_table,
                                                 batch_ids,
                                                 tile_ids_per_batch,
@@ -265,10 +264,9 @@ void CascadeAppendAttentionKernel(
                                                 causal,
                                                 is_decoder,
                                                 enable_prefill,
-                                                cache_quant_type_str,
                                                 stream,
                                                 out);
-    } else if (cache_quant_type_str == "cache_fp8" or cache_quant_type_str == "block_wise_fp8") {
+    } else if (cache_quant_type_str == "cache_fp8") {
         CascadeAppendAttentionC8Kernel<T, OutT, true>(meta_data,
                                                 qkv,
                                                 cache_k,
@@ -283,8 +281,8 @@ void CascadeAppendAttentionKernel(
                                                 seq_lens_q,
                                                 seq_lens_kv,
                                                 seq_lens_encoder,
-                                                batch_id_per_token,
-                                                cu_seqlens_q,
+                                                padding_offsets,
+                                                cum_offsets,
                                                 block_table,
                                                 batch_ids,
                                                 tile_ids_per_batch,
@@ -301,7 +299,6 @@ void CascadeAppendAttentionKernel(
                                                 causal,
                                                 is_decoder,
                                                 enable_prefill,
-                                                cache_quant_type_str,
                                                 stream,
                                                 out);
     } else if (cache_quant_type_str == "cache_int4_zp") {
@@ -319,8 +316,8 @@ void CascadeAppendAttentionKernel(
                                                 seq_lens_q,
                                                 seq_lens_kv,
                                                 seq_lens_encoder,
-                                                batch_id_per_token,
-                                                cu_seqlens_q,
+                                                padding_offsets,
+                                                cum_offsets,
                                                 block_table,
                                                 batch_ids,
                                                 tile_ids_per_batch,

custom_ops/gpu_ops/append_attn/decode_attention_func.cuh

@@ -1,236 +0,0 @@
-// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-#pragma once
-
-
-#include "multi_head_latent_attention_kernel.h"
-
-template <size_t vec_size, typename T>
-struct softmax_state_t {
-  AlignedVector<T, vec_size> o;
-  T m;
-  T d;
-
-  __device__ __forceinline__ void init() {
-    if constexpr (std::is_same<T, half>::value) {
-#pragma unroll
-      for (int i = 0; i < vec_size / 2; ++i) {
-        *((half2*)(&o) + i) = make_half2(0, 0);
-      }
-    } else if constexpr (std::is_same<T, __nv_bfloat16>::value) {
-#pragma unroll
-      for (int i = 0; i < vec_size / 2; ++i) {
-        *((nv_bfloat162*)(&o) + i) = make_bfloat162(0, 0);
-      }
-    }
-    d = 1.f;
-    if constexpr (std::is_same<T, half>::value) {
-      m = __float2half(-5e4f);
-    } else if constexpr (std::is_same<T, nv_bfloat16>::value) {
-      m = __float2bfloat16(-3.38953e38f);
-    }
-  }
-
-  __device__ __forceinline__ softmax_state_t() {
-    init();
-  }
-
-  __device__ __forceinline__ void merge(const AlignedVector<T, vec_size>& other_o,
-                                        T other_m,
-                                        T other_d) {
-    // using kType = typename cascade_attn_nv_type2_traits<T>::type;
-    T m_prev = m, d_prev = d;
-    m = m_prev > other_m ? m_prev : other_m;
-    T scale1 = hexp(m_prev - m), scale2 = hexp(other_m - m);
-
-    d = d_prev * scale1 + other_d * scale2;
-
-#pragma unroll
-    for (size_t i = 0; i < vec_size; ++i) {
-      o[i] = o[i] * scale1 + other_o[i] * scale2;
-    }
-  }
-
-  __device__ __forceinline__ void normalize() {
-
-#pragma unroll
-    for (size_t i = 0; i < vec_size; ++i) {
-      o[i] /= d;
-    }
-  }
-
-};
-
-template <size_t vec_size, typename T, uint32_t num_tiles = 0>
-struct softmax_state_ts {
-  uint32_t num_tiles_ = num_tiles;
-  AlignedVector<T, vec_size> o[num_tiles];
-  float m;
-  float d;
-
-  __device__ __forceinline__ void init() {
-#pragma unroll
-    for (uint32_t tile_id = 0; tile_id < num_tiles_; ++tile_id) {
-      if constexpr (std::is_same<T, half>::value) {
-#pragma unroll
-        for (int i = 0; i < vec_size / 2; ++i) {
-          *((half2*)(&o[tile_id]) + i) = make_half2(0, 0);
-        }
-      } else if constexpr (std::is_same<T, __nv_bfloat16>::value) {
-#pragma unroll
-        for (int i = 0; i < vec_size / 2; ++i) {
-          *((nv_bfloat162*)(&o[tile_id]) + i) = make_bfloat162(0, 0);
-        }
-      }
-    }
-    d = 1.f;
-    if constexpr (std::is_same<T, half>::value) {
-      m = -5e4f;
-    } else if constexpr (std::is_same<T, nv_bfloat16>::value) {
-      m = -3.38953e38f;
-    }
-  }
-
-  __device__ __forceinline__ softmax_state_ts() {
-    init();
-  }
-
-  __device__ __forceinline__ void normalize(const uint32_t tile_id) {
-
-#pragma unroll
-    for (size_t i = 0; i < vec_size; i++) {
-      o[tile_id][i] /= d;
-    }
-  }
-
-};
-
-template <SharedMemFillMode fill_mode, uint32_t HEAD_DIM_QK, uint32_t vec_size, uint32_t NUM_VEC_PER_HEAD, uint32_t bdx, uint32_t BLOCK_SIZE, uint32_t CACHE_VEC_SIZE, typename CacheT>
-__device__ __forceinline__ void produce_kv(CacheT *smem,
-                                          CacheT *kv_base_gptr,
-                                          const int * block_table_smem,
-                                          const uint32_t seq_offset_gmem,
-                                          const uint32_t seq_offset_smem,
-                                          const uint32_t kv_head_idx,
-                                          const uint32_t kv_num_heads,
-                                          const uint32_t tidx,
-                                          const uint32_t chunk_start,
-                                          const uint32_t chunk_end) {
-  int block_id = __ldg(&block_table_smem[seq_offset_gmem / BLOCK_SIZE]);
-  if (block_id < 0) {
-    block_id = 0;
-  }
-  const uint32_t block_offset = seq_offset_gmem % BLOCK_SIZE;
-  // 8/16 T/int8 each time
-  const uint32_t k_offset_base = ((block_id * kv_num_heads + kv_head_idx) * BLOCK_SIZE + block_offset) * HEAD_DIM_QK;
-  const uint32_t smem_offset_base = seq_offset_smem * HEAD_DIM_QK;
-  for(uint32_t vid = tidx; vid < NUM_VEC_PER_HEAD; vid += bdx) {
-    pred_load<128, PrefetchMode::kPrefetch, fill_mode, CacheT>(
-      smem + smem_offset_base + vid * CACHE_VEC_SIZE,
-      kv_base_gptr + k_offset_base + vid * CACHE_VEC_SIZE,
-      seq_offset_gmem < chunk_end
-    );
-  }
-}
-
-template <uint32_t vec_size, uint32_t NUM_VEC_PER_HEAD, uint32_t bdx, uint32_t bdy, uint32_t HEAD_DIM, uint32_t DEAL_EACH_TIME, uint32_t num_tile_v, typename T, typename CacheT>
-__device__ __forceinline__ void compute_qk(const T* cu_q_smem,
-                                           const CacheT* k_smem,
-                                           const uint32_t kv_idx_base,
-                                           const uint32_t stage_idx,
-                                           const uint32_t iter_base,
-                                           const uint32_t iter_bound,
-                                           const uint32_t tidx,
-                                           const uint32_t gid,
-                                           const float scale,
-                                           float *s,
-                                           softmax_state_ts<vec_size, T, num_tile_v>& st) {
-  const CacheT* smem;
-  AlignedVector<T, vec_size> q_vec;
-  AlignedVector<T, vec_size> k_vec;
-  float m_prev = st.m;
-  // smem = base_smem + (stage_idx * DEAL_EACH_TIME + zid * tile_size) * HEAD_DIM;
-  smem = k_smem + stage_idx * DEAL_EACH_TIME * HEAD_DIM;
-#pragma unroll
-  for (uint32_t j = 0; j < DEAL_EACH_TIME; ++j) {
-    if (iter_base + j < iter_bound) {
-      if constexpr (std::is_same<T, half>::value) {
-        s[j] = 0.f;
-      } else if constexpr (std::is_same<T, __nv_bfloat16>::value) {
-        s[j] = 0.f;
-      }
-#pragma unroll
-      for(uint32_t vid = tidx; vid < NUM_VEC_PER_HEAD; vid += bdx) {
-        Load<T, vec_size>(cu_q_smem + vid * vec_size, &q_vec);
-        Load<CacheT, vec_size>(smem + j * HEAD_DIM + vid * vec_size, &k_vec);
-        for (uint32_t i = 0; i < vec_size; ++i) {
-          s[j] += static_cast<float>(q_vec[i] * k_vec[i]);
-        }
-      }
-#pragma unroll
-      for (uint32_t offset = bdx / 2; offset > 0; offset /= 2) {
-        s[j] += __shfl_xor_sync(-1, s[j], offset, 32);
-      }
-      __syncthreads();
-    } else {
-      if constexpr (std::is_same<T, half>::value) {
-        s[j] = -5e4f;
-      } else if constexpr (std::is_same<T, __nv_bfloat16>::value) {
-        s[j] = -3.38953e38f;
-      }
-    }
-    st.m = st.m > s[j] ? st.m : s[j];
-  }
-
-  // T o_scale = hexp(m_prev - st.m);
-  float o_scale = __expf(m_prev - st.m);
-  st.d *= o_scale;
-
-#pragma unroll
-  for (uint32_t j = 0; j < DEAL_EACH_TIME; ++j) {
-    // s[j] = hexp(s[j] - st.m);
-    s[j] = __expf(s[j] - st.m);
-    st.d += s[j];
-  }
-#pragma unroll
-  for (uint32_t tile_id = 0; tile_id < num_tile_v; ++tile_id) {
-    for (uint32_t i = 0; i < vec_size; ++i) {
-      st.o[tile_id][i] *= o_scale;
-    }
-  }
-}
-
-template<uint32_t vec_size, uint32_t NUM_VEC_PER_HEAD, uint32_t bdx, uint32_t DEAL_EACH_TIME, uint32_t HEAD_DIM_QK, uint32_t num_tile, typename T, typename CacheT>
-__device__ __forceinline__ void compute_sv(const float *s,
-                                           const CacheT *base_v_smem,
-                                           const uint32_t stage_idx,
-                                           const uint32_t iter_base,
-                                           const uint32_t iter_bound,
-                                           const uint32_t tidx,
-                                           softmax_state_ts<vec_size, T, num_tile>& st) {
-  const CacheT* v_smem;
-  AlignedVector<T, vec_size> v_vec;
-#pragma unroll
-  for (int j = 0; (j < DEAL_EACH_TIME) && (iter_base + j < iter_bound); ++j) {
-    v_smem = base_v_smem + stage_idx * DEAL_EACH_TIME * HEAD_DIM_QK + j * HEAD_DIM_QK;
-    for(uint32_t vid = tidx; vid < NUM_VEC_PER_HEAD; vid += bdx) {
-      Load<T, vec_size>(v_smem + vid * vec_size, &v_vec);
-      uint32_t tile_id = vid / bdx;
-#pragma unroll
-      for (int reg_id = 0; reg_id < vec_size; ++reg_id) {
-        st.o[tile_id][reg_id] += static_cast<T>(s[j]) * v_vec[reg_id];
-      }
-    }
-  }
-}

custom_ops/gpu_ops/append_attn/decode_attention_kernel.cu

@@ -1,560 +0,0 @@
-// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#include "decode_attention_func.cuh"
-
-#define CHECK(call)                                                             \
-do                                                                              \
-{                                                                               \
-    const cudaError_t error_code = call;                                        \
-    if (error_code != cudaSuccess)                                              \
-    {                                                                           \
-        printf("CUDA Error:\n");                                                \
-        printf("     File:      %s\n", __FILE__);                               \
-        printf("     Line       %d:\n", __LINE__);                              \
-        printf("     Error code:%d\n", error_code);                             \
-        printf("     Error text:%s\n", cudaGetErrorString(error_code));         \
-        exit(1);                                                                \
-    }                                                                           \
-}while(0)
-
-template <typename T, typename OutT, int vec_size, uint32_t bdy, uint32_t HEAD_DIM>
-__global__ void merge_varlen_multi_chunks_v2_kernel(const T * __restrict__ multi_out, // [bsz, num_chunks, num_heads, head_dim]
-                                                    const T * __restrict__ multi_m, // [bsz, num_chunks, num_heads]
-                                                    const T * __restrict__ multi_d, // [bsz, num_chunks, num_heads]
-                                                    const int * __restrict__ seq_lens_q,
-                                                    const int * __restrict__ seq_lens_kv,
-                                                    const int * __restrict__ cu_seqlens_q,
-                                                    const T * __restrict__ shift_bias, // [q_num_heads * HEAD_DIM]
-                                                    const T * __restrict__ smooth_weight, // [q_num_heads * HEAD_DIM]
-                                                    OutT * __restrict__ out, // [token_num, num_heads, head_dim]
-                                                    const float in_scale,
-                                                    const int num_chunks,
-                                                    const int chunk_size,
-                                                    const int max_seq_len,
-                                                    const int num_heads,
-                                                    const int head_dim) {
-  const int vid = threadIdx.x, ty = threadIdx.y;
-  const int qid = blockIdx.x, hid = blockIdx.y;
-  const int seq_len_q = seq_lens_q[qid];
-  if (seq_len_q == 0) return;
-  int seq_len_kv = seq_lens_kv[qid];
-  if (seq_len_kv == 0) return;
-  seq_len_kv += seq_len_q;
-  const int num_chunks_this_seq = div_up(seq_len_kv, chunk_size);
-  if (num_chunks_this_seq == 1 || ty >= num_chunks_this_seq) {
-    return;
-  }
-  __shared__ T smem[bdy * HEAD_DIM];
-  __shared__ T md_smem[bdy * 2];
-
-  const int start_token_ids = cu_seqlens_q[qid];
-  using LoadT = AlignedVector<T, vec_size>;
-  LoadT load_vec;
-  LoadT res_vec;
-  if constexpr (std::is_same<T, half>::value) {
-#pragma unroll
-    for (int i = 0; i < vec_size / 2; ++i) {
-      *((half2*)(&res_vec) + i) = make_half2(0, 0);
-    }
-  } else if constexpr (std::is_same<T, nv_bfloat16>::value) {
-#pragma unroll
-    for (int i = 0; i < vec_size / 2; ++i) {
-      *((nv_bfloat162*)(&res_vec) + i) = make_bfloat162(0, 0);
-    }
-  }
-  T m;
-  T d = 1.f;
-  if constexpr (std::is_same<T, half>::value) {
-    m = __float2half(-5e4f);
-  } else if constexpr (std::is_same<T, nv_bfloat16>::value) {
-    m = __float2bfloat16(-3.38953e38f);
-  }
-  // merge per ty
-#pragma unroll 2
-  for (int i = ty; i < num_chunks_this_seq; i += bdy) {
-    uint32_t offset = (qid * num_chunks + i) * num_heads + hid;
-    T m_prev = m;
-    T d_prev = d;
-    const T m_now = multi_m[offset];
-    const T d_now = multi_d[offset];
-    m = m_prev > m_now ? m_prev : m_now;
-    offset = (qid * num_chunks * num_heads + i * num_heads + hid) * head_dim + vid * vec_size;
-    Load<T, vec_size>(&multi_out[offset], &load_vec);
-    const T scale1 = hexp(m_prev - m), scale2 = hexp(m_now - m);
-    d = d * scale1 + d_now * scale2;
-#pragma once
-    for (int j = 0; j < vec_size; j++) {
-      res_vec[j] = res_vec[j] * scale1 + load_vec[j] * scale2;
-    }
-  }
-  // store ty res
-  Store<T, vec_size>(res_vec, &smem[ty * head_dim + vid * vec_size]);
-  md_smem[2 * ty] = m;
-  md_smem[2 * ty + 1] = d;
-  __syncthreads();
-
-  // merge bdy
-  softmax_state_t<vec_size, T> st{};
-  const uint32_t iter_num = min(num_chunks_this_seq, bdy);
-#pragma once
-  for (int i = 0; i < iter_num; i++) {
-    Load<T, vec_size>(&smem[i * head_dim + vid * vec_size], &load_vec);
-    const T m_tmp = md_smem[2 * i], d_tmp = md_smem[2 * i + 1];
-    st.merge(load_vec, m_tmp, d_tmp);
-  }
-  st.normalize();
-
-  AlignedVector<OutT, vec_size> out_vec;
-
-#pragma unroll
-  for (int i = 0; i < vec_size; ++i) {
-    out_vec[i] = static_cast<OutT>(st.o[i]);
-  }
-  Store<OutT, vec_size>(out_vec, &out[(start_token_ids * num_heads + hid) * head_dim + vid * vec_size]);
-}
-
-template <bool partition_kv, typename T, typename OutT, typename CacheT, uint32_t NUM_STAGES, uint32_t DEAL_EACH_TIME, uint32_t GROUP_SIZE, uint32_t HEAD_DIM_QK, uint32_t HEAD_DIM_V,
-          uint32_t BLOCK_SIZE, uint32_t VEC_SIZE, uint32_t CACHE_VEC_SIZE, uint32_t bdx, uint32_t bdy>
-__global__ void multi_query_decode_attention_kernel(T * __restrict__ q, // [token_num, num_heads, head_dim]
-                                                    CacheT * __restrict__ cache_k, // [max_block_num, num_heads, block_size, head_dim]
-                                                    CacheT * __restrict__ cache_v,
-                                                    const T * __restrict__ shift_bias, // [q_num_heads * HEAD_DIM]
-                                                    const T * __restrict__ smooth_weight, // [q_num_heads * HEAD_DIM]
-                                                    const int * __restrict__ seq_lens_q,
-                                                    const int * __restrict__ seq_lens_kv,
-                                                    const int * __restrict__ cu_seqlens_q,
-                                                    const int * __restrict__ block_table, // [bsz, block_num_per_seq]
-                                                    const int max_seq_len,
-                                                    const int max_dec_len,
-                                                    const int max_block_num_per_seq,
-                                                    const float scale,
-                                                    const float in_scale,
-                                                    const uint32_t chunk_size,
-                                                    T * __restrict__ tmp_workspace, // [batch_size, num_chunks, num_heads, head_dim]
-                                                    T * __restrict__ tmp_m, // [batch_size, num_chunks, num_heads]
-                                                    T * __restrict__ tmp_d, // [batch_size, num_chunks, num_heads]
-                                                    OutT * __restrict__ out) {
-  const uint32_t bidx = blockIdx.x, kv_head_idx = blockIdx.z;
-  const uint32_t bid = bidx, gid = threadIdx.y;
-  const uint32_t tidx = threadIdx.x;
-  constexpr uint32_t num_vec_per_head_qk = HEAD_DIM_QK / VEC_SIZE;
-  constexpr uint32_t num_vec_per_head_v = HEAD_DIM_V / VEC_SIZE;
-  constexpr uint32_t num_tile_v = (num_vec_per_head_v + bdx - 1) / bdx;
-
-  const uint32_t q_head_idx = kv_head_idx * GROUP_SIZE + gid;
-  const uint32_t kv_num_heads = gridDim.z;
-  const uint32_t q_num_heads = kv_num_heads * GROUP_SIZE;
-
-  const int *block_table_now = block_table + bid * max_block_num_per_seq;
-
-  const uint32_t num_chunks = gridDim.y;
-  const uint32_t chunk_id = blockIdx.y;
-  const uint32_t q_len = seq_lens_q[bid];
-  if (q_len <= 0) {
-    return;
-  }
-  uint32_t kv_len = seq_lens_kv[bid]; // !!!!!!!!
-  if (kv_len <= 0) {
-     return;
-  }
-  kv_len += q_len;
-  const uint32_t num_chunk_this_seq = div_up(kv_len, chunk_size);
-  const uint32_t q_start_idx = cu_seqlens_q[bid];
-  const uint32_t q_write_idx = cu_seqlens_q[bid];
-  if (chunk_id >= num_chunk_this_seq) {
-    return;
-  }
-
-  const uint32_t chunk_start = partition_kv ? chunk_id * chunk_size : 0;
-  const uint32_t chunk_end = partition_kv ? min(kv_len, chunk_start + chunk_size) : kv_len;
-  const uint32_t chunk_len = chunk_end - chunk_start;
-
-  extern __shared__ uint8_t smem[];
-  const T *q_now = q + (q_start_idx * q_num_heads + q_head_idx) * HEAD_DIM_QK;
-  T *q_smem = reinterpret_cast<T*>(smem); // [HEAD_DIM_QK * sizeof(T)]
-  T *cu_q_smem = q_smem + gid * HEAD_DIM_QK;
-#pragma unroll
-  for(uint32_t vid = tidx; vid < num_vec_per_head_qk; vid += bdx) {
-    ((float4*)(&cu_q_smem[vid * VEC_SIZE]))[0] = ((float4*)(&q_now[vid * VEC_SIZE]))[0];
-
-  }
-  __syncthreads();
-  using VecT = AlignedVector<T, VEC_SIZE>;
-  VecT q_vec;
-#pragma unroll
-  for(uint32_t vid = tidx; vid < num_vec_per_head_qk; vid += bdx) {
-    Load<T, VEC_SIZE>(cu_q_smem + vid * VEC_SIZE, &q_vec);
-    for (uint32_t i = 0; i < VEC_SIZE; ++i) {
-      q_vec[i] *= scale;
-    }
-    Store<T, VEC_SIZE>(q_vec, cu_q_smem + vid * VEC_SIZE);
-  }
-
-
-  CacheT *kv_smem = reinterpret_cast<CacheT*>(smem + GROUP_SIZE * HEAD_DIM_QK * sizeof(CacheT));
-  uint32_t stage_idx = 0;
-  constexpr int loop_times = DEAL_EACH_TIME / bdy;
-#pragma unroll
-  for (int i = 0; i < NUM_STAGES; ++i) {
-#pragma unroll
-    for (int j = 0; j < loop_times; ++j) {
-      const uint32_t k_seq_offset = i * DEAL_EACH_TIME + j * bdy + gid;
-      const uint32_t k_seq_id = chunk_start + k_seq_offset;
-      produce_kv<SharedMemFillMode::kNoFill, HEAD_DIM_QK, VEC_SIZE, num_vec_per_head_qk, bdx, BLOCK_SIZE, CACHE_VEC_SIZE>(
-        kv_smem,
-        cache_k,
-        block_table_now,
-        k_seq_id,
-        k_seq_offset,
-        kv_head_idx,
-        kv_num_heads,
-        tidx,
-        chunk_start,
-        chunk_end
-      );
-    }
-    commit_group();
-    stage_idx = (stage_idx + 1) % NUM_STAGES;
-  }
-
-
-  softmax_state_ts<VEC_SIZE, T, num_tile_v> st;
-  float s[DEAL_EACH_TIME];
-
-  const uint32_t num_iters = div_up(chunk_len, DEAL_EACH_TIME);
-  for (int iter = 0; iter < num_iters; ++iter) {
-    wait_group<NUM_STAGES - 1>();
-    __syncthreads();
-    // compute qk
-    compute_qk<VEC_SIZE, num_vec_per_head_qk, bdx, bdy, HEAD_DIM_QK, DEAL_EACH_TIME, num_tile_v>(
-      cu_q_smem,
-      kv_smem,
-      chunk_start + iter * DEAL_EACH_TIME,
-      stage_idx,
-      iter * DEAL_EACH_TIME,
-      chunk_len,
-      tidx,
-      gid,
-      scale,
-      s,
-      st
-    );
-    __syncthreads();
-
-    // compute sv
-    compute_sv<VEC_SIZE, num_vec_per_head_v, bdx, DEAL_EACH_TIME, HEAD_DIM_QK, num_tile_v>(
-      s,
-      kv_smem,
-      stage_idx,
-      iter * DEAL_EACH_TIME,
-      chunk_len,
-      tidx,
-      st
-    );
-    __syncthreads();
-
-#pragma unroll
-    for (int j = 0; j < loop_times; ++j) {
-      const uint32_t k_seq_offset = j * bdy + gid;
-      produce_kv<SharedMemFillMode::kNoFill, HEAD_DIM_QK, VEC_SIZE, num_vec_per_head_qk, bdx, BLOCK_SIZE, CACHE_VEC_SIZE>(
-        kv_smem,
-        cache_k,
-        block_table_now,
-        chunk_start + k_seq_offset + (iter + NUM_STAGES) * DEAL_EACH_TIME,
-        stage_idx * DEAL_EACH_TIME + k_seq_offset,
-        kv_head_idx,
-        kv_num_heads,
-        tidx,
-        chunk_start,
-        chunk_end
-      );
-    }
-    commit_group();
-    stage_idx = (stage_idx + 1) % NUM_STAGES;
-  }
-  wait_group<0>();
-  __syncthreads();
-
-  // normize if not partition_kv
-  for(uint32_t vid = tidx; vid < num_vec_per_head_v; vid += bdx) {
-    const uint32_t tile_id = vid / bdx;
-    if (!partition_kv || num_chunk_this_seq == 1) {
-      st.normalize(tile_id);
-    }
-    if (partition_kv && num_chunk_this_seq > 1) {
-      const uint32_t head_idx = (bid * num_chunks + chunk_id) * q_num_heads + q_head_idx;
-      Store<T, VEC_SIZE>(st.o[tile_id], tmp_workspace + head_idx * HEAD_DIM_V + vid * VEC_SIZE);
-      tmp_m[head_idx] = st.m;
-      tmp_d[head_idx] = st.d;
-    } else {
-      Store<OutT, VEC_SIZE>(st.o[tile_id], out + (q_write_idx * q_num_heads + q_head_idx) * HEAD_DIM_V + vid * VEC_SIZE);
-    }
-  }
-}
-
-
-template <typename T, uint32_t GROUP_SIZE, uint32_t HEAD_DIM_QK, uint32_t HEAD_DIM_V, uint32_t BLOCK_SIZE, bool CAUSAL, uint32_t NUM_STAGE, uint32_t cache_bytes, uint32_t DEAL_EACH_TIME>
-void MultiQueryDecoderAttention(
-  const AppendAttnMetaData& meta_data,
-  cudaStream_t &stream,
-  const paddle::Tensor &q,
-  const paddle::Tensor &cache_k, // [max_block_num, num_kv_heads, block_size, head_dim]
-  const paddle::Tensor &cache_v, // [num_kv_heads, head_dim]
-  const paddle::optional<paddle::Tensor>& attn_mask,
-  const paddle::optional<paddle::Tensor>& shift_bias,
-  const paddle::optional<paddle::Tensor>& smooth_weight,
-  const paddle::Tensor &seq_lens_q,
-  const paddle::Tensor &seq_lens_kv,
-  const paddle::Tensor &batch_id_per_token,
-  const paddle::Tensor &cu_seqlens_q,
-  const paddle::Tensor &block_table,
-  const int max_seq_len,
-  const int max_dec_len,
-  const float rope_scale,
-  const float rope_theta,
-  const float softmax_scale,
-  const float in_scale,
-  paddle::Tensor *out) {
-  using NV_TYPE = typename cascade_attn_type_traits<T>::type;
-
-  auto num_heads = meta_data.q_num_heads;
-  auto kv_num_heads = meta_data.kv_num_heads;
-  auto token_num = meta_data.token_nums;
-  auto bsz = meta_data.batch_size;
-  auto max_block_num_per_seq = meta_data.max_blocks_per_seq;
-  constexpr int num_stages = NUM_STAGE;
-
-  constexpr int vec_size = 16 / sizeof(T); // 8 16 32
-  constexpr int cache_vec_size = 128 / cache_bytes; // 8 16 32
-  constexpr int blockxc = HEAD_DIM_QK / cache_vec_size;
-  constexpr int num_vec_per_head = HEAD_DIM_QK / vec_size;
-  constexpr int blockx = num_vec_per_head < 32 ? num_vec_per_head : 32;
-
-  constexpr int blocky = GROUP_SIZE;
-  const int gridx = bsz;
-
-  constexpr int num_threads = blockx * blocky;
-
-  auto splitkv_kernel = multi_query_decode_attention_kernel<true, NV_TYPE, NV_TYPE, NV_TYPE, num_stages, DEAL_EACH_TIME, GROUP_SIZE, HEAD_DIM_QK, HEAD_DIM_V,
-                                                                        BLOCK_SIZE, vec_size, cache_vec_size, blockx, blocky>;
-  uint32_t cache_smem_bytes = 0;
-
-  const T *shift_bias_ptr = shift_bias ? shift_bias.get().data<T>() : nullptr;
-  const T *smooth_weight_ptr = smooth_weight ? smooth_weight.get().data<T>() : nullptr;
-  cache_smem_bytes = num_stages * DEAL_EACH_TIME * HEAD_DIM_QK * sizeof(T);
-
-  const uint32_t chunk_size = get_max_partition_size(bsz);
-  const int num_chunks = div_up(max_dec_len, chunk_size);
-  size_t smem_size = cache_smem_bytes + GROUP_SIZE * HEAD_DIM_QK * sizeof(T);
-
-  if (smem_size >= 48 * 1024) {
-    cudaFuncSetAttribute(
-      splitkv_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
-  }
-  const int dev_id = 0;
-  int sm_count;
-  int act_blocks_per_sm;
-  cudaDeviceGetAttribute(&sm_count, cudaDevAttrMultiProcessorCount, dev_id);
-  cudaOccupancyMaxActiveBlocksPerMultiprocessor(
-    &act_blocks_per_sm, splitkv_kernel, num_threads, smem_size);
-  assert(act_blocks_per_sm > 1);
-
-  const int num_blocks_per_wave = sm_count * act_blocks_per_sm;
-  const int num_blocks_need = gridx * num_chunks * kv_num_heads;
-  const int max_num_chunks = div_up(num_blocks_per_wave, num_blocks_need);
-  const float ratio = static_cast<float>(num_blocks_need) / static_cast<float>(num_blocks_per_wave);
-
-  dim3 grids(gridx, num_chunks, kv_num_heads);
-  dim3 blocks(blockx, blocky);
-  if (num_chunks <= 1) {
-    auto no_splitkv_kernel = multi_query_decode_attention_kernel<false, NV_TYPE, NV_TYPE, NV_TYPE, num_stages, DEAL_EACH_TIME, GROUP_SIZE, HEAD_DIM_QK, HEAD_DIM_V, BLOCK_SIZE, vec_size,
-                                                                             cache_vec_size, blockx, blocky>;
-    if (smem_size >= 48 * 1024) {
-      cudaFuncSetAttribute(
-        no_splitkv_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size);
-    }
-    no_splitkv_kernel<<<grids, blocks, smem_size, stream>>>(
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(q.data<T>())),
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(cache_k.data<T>())),
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(cache_v.data<T>())),
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(shift_bias_ptr)),
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(smooth_weight_ptr)),
-      seq_lens_q.data<int>(),
-      seq_lens_kv.data<int>(),
-      cu_seqlens_q.data<int>(),
-      block_table.data<int>(),
-      max_seq_len,
-      max_dec_len,
-      max_block_num_per_seq,
-      softmax_scale,
-      in_scale,
-      chunk_size,
-      nullptr,
-      nullptr,
-      nullptr,
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(out->data<T>()))
-    );
-
-    // CHECK(cudaGetLastError());
-    // CHECK(cudaDeviceSynchronize());
-  } else {
-    auto *allocator = paddle::GetAllocator(q.place());
-    phi::Allocator::AllocationPtr tmp_workspace, tmp_m, tmp_d;
-    tmp_workspace = allocator->Allocate(
-        phi::SizeOf(q.dtype()) *
-        static_cast<size_t>(bsz * num_chunks * num_heads * HEAD_DIM_V));
-    tmp_m = allocator->Allocate(
-        phi::SizeOf(q.dtype()) *
-        static_cast<size_t>(bsz * num_chunks * num_heads));
-    tmp_d = allocator->Allocate(
-        phi::SizeOf(q.dtype()) *
-        static_cast<size_t>(bsz * num_chunks * num_heads));
-
-    splitkv_kernel<<<grids, blocks, smem_size, stream>>>(
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(q.data<T>())),
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(cache_k.data<T>())),
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(cache_v.data<T>())),
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(shift_bias_ptr)),
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(smooth_weight_ptr)),
-      seq_lens_q.data<int>(),
-      seq_lens_kv.data<int>(),
-      cu_seqlens_q.data<int>(),
-      block_table.data<int>(),
-      max_seq_len,
-      max_dec_len,
-      max_block_num_per_seq,
-      softmax_scale,
-      in_scale,
-      chunk_size,
-      reinterpret_cast<NV_TYPE*>(tmp_workspace->ptr()),
-      reinterpret_cast<NV_TYPE*>(tmp_m->ptr()),
-      reinterpret_cast<NV_TYPE*>(tmp_d->ptr()),
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(out->data<T>()))
-    );
-    // CHECK(cudaGetLastError());
-    // CHECK(cudaDeviceSynchronize());
-
-    constexpr int mblockx = HEAD_DIM_V / vec_size;
-    constexpr int bdy = 256 / mblockx;
-    dim3 grids_merge(bsz, num_heads);
-    dim3 blocks_merge(mblockx, bdy);
-    merge_varlen_multi_chunks_v2_kernel<NV_TYPE, NV_TYPE, vec_size, bdy, HEAD_DIM_V><<<grids_merge, blocks_merge, 0, stream>>>(
-      reinterpret_cast<NV_TYPE*>(tmp_workspace->ptr()),
-      reinterpret_cast<NV_TYPE*>(tmp_m->ptr()),
-      reinterpret_cast<NV_TYPE*>(tmp_d->ptr()),
-      seq_lens_q.data<int>(),
-      seq_lens_kv.data<int>(),
-      cu_seqlens_q.data<int>(),
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(shift_bias_ptr)),
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(smooth_weight_ptr)),
-      reinterpret_cast<NV_TYPE*>(const_cast<T*>(out->data<T>())),
-      in_scale,
-      num_chunks,
-      chunk_size,
-      max_seq_len,
-      num_heads,
-      HEAD_DIM_V
-    );
-  }
-  // CHECK(cudaGetLastError());
-  // CHECK(cudaDeviceSynchronize());
-}
-
-template <typename T>
-void DecodeMLAAttentionKernel(
-  const AppendAttnMetaData& meta_data,
-  const paddle::Tensor &q, // [token_num, num_heads, head_dim]
-  const paddle::Tensor &cache_k,
-  const paddle::Tensor &cache_v,
-  const paddle::optional<paddle::Tensor>& attn_mask,
-  const paddle::optional<paddle::Tensor>& shift_bias,
-  const paddle::optional<paddle::Tensor>& smooth_weight,
-  const paddle::Tensor &seq_lens_q, // q_seq_len is 1
-  const paddle::Tensor &seq_lens_kv,
-  const paddle::Tensor &batch_id_per_token,
-  const paddle::Tensor &cu_seqlens_q,
-  const paddle::Tensor &block_table,
-  int max_seq_len,
-  int max_dec_len,
-  float softmax_scale,
-  float in_scale,
-  bool causal,
-  cudaStream_t &stream,
-  paddle::Tensor *out) {
-  const auto token_num = meta_data.token_nums;
-  const auto block_size = meta_data.block_size;
-  const auto bsz = meta_data.batch_size;
-  const auto num_heads = meta_data.q_num_heads;
-  const auto group_size = meta_data.q_num_heads / meta_data.kv_num_heads;
-  const auto head_dim_qk = meta_data.head_dims;
-  const auto head_dim_v = meta_data.head_dims_v;
-  const float rope_scale = 0.0;
-  const float rope_theta = 0.0;
-  const uint32_t deal_each_time = get_cascade_attention_deal_each_time();
-  const uint32_t num_stage = get_cascade_attention_num_stages();
-  const uint32_t num_threads = get_cascade_attention_num_threads();
-
-  DISPATCH_CAUSAL(causal, CAUSAL,
-    {DISPATCH_MLA_GROUP_SIZE(group_size, GROUP_SIZE,
-      {DISPATCH_MLA_HEAD_DIM(head_dim_qk, HEAD_DIM_QK,
-        {DISPATCH_MLA_HEAD_DIM(head_dim_v, HEAD_DIM_V,
-          {DISPATCH_BLOCK_SIZE(block_size, BLOCK_SIZE,
-              {DISPATCH_DEAL_EACH_TIME(deal_each_time, DEAL_EACH_TIME,
-                  {MultiQueryDecoderAttention<T, GROUP_SIZE, HEAD_DIM_QK, HEAD_DIM_V, BLOCK_SIZE, CAUSAL, 2, 16, DEAL_EACH_TIME>(
-                  meta_data, stream, q, cache_k, cache_v, attn_mask, shift_bias, smooth_weight, seq_lens_q, seq_lens_kv, batch_id_per_token, cu_seqlens_q,
-                  block_table, max_seq_len, max_dec_len, rope_scale, rope_theta, softmax_scale, in_scale, out);})})})})})});
-}
-
-template void DecodeMLAAttentionKernel<paddle::bfloat16>(
-  const AppendAttnMetaData& meta_data,
-  const paddle::Tensor &q, // [token_num, num_heads, head_dim]
-  const paddle::Tensor &cache_k,
-  const paddle::Tensor &cache_v,
-  const paddle::optional<paddle::Tensor>& attn_mask,
-  const paddle::optional<paddle::Tensor>& shift_bias,
-  const paddle::optional<paddle::Tensor>& smooth_weight,
-  const paddle::Tensor &seq_lens_q, // q_seq_len is 1
-  const paddle::Tensor &seq_lens_kv,
-  const paddle::Tensor &batch_id_per_token,
-  const paddle::Tensor &cu_seqlens_q,
-  const paddle::Tensor &block_table,
-  int max_seq_len,
-  int max_dec_len,
-  float softmax_scale,
-  float in_scale,
-  bool causal,
-  cudaStream_t &stream,
-  paddle::Tensor *out);
-
-template void DecodeMLAAttentionKernel<paddle::float16>(
-  const AppendAttnMetaData& meta_data,
-  const paddle::Tensor &q, // [token_num, num_heads, head_dim]
-  const paddle::Tensor &cache_k,
-  const paddle::Tensor &cache_v,
-  const paddle::optional<paddle::Tensor>& attn_mask,
-  const paddle::optional<paddle::Tensor>& shift_bias,
-  const paddle::optional<paddle::Tensor>& smooth_weight,
-  const paddle::Tensor &seq_lens_q, // q_seq_len is 1
-  const paddle::Tensor &seq_lens_kv,
-  const paddle::Tensor &batch_id_per_token,
-  const paddle::Tensor &cu_seqlens_q,
-  const paddle::Tensor &block_table,
-  int max_seq_len,
-  int max_dec_len,
- float softmax_scale,
-  float in_scale,
-  bool causal,
-  cudaStream_t &stream,
-  paddle::Tensor *out);

custom_ops/gpu_ops/append_attn/decoder_write_cache_with_rope_impl.cuh

@@ -18,142 +18,6 @@
 #include "mma_tensor_op.cuh"
 #include "utils.cuh"
 
-template <typename T, int VecSize = 1>
-__global__ void append_decode_cache_T_rope_qk_norm_kernel(
-    const T* __restrict__ quant_qkv,  // [bsz, num_heads + 2 * kv_num_heads,
-                                      // head_size]
-    T* __restrict__ key_cache,    // [num_blocks, kv_num_heads, block_size,
-                                  // head_size // 2]
-    T* __restrict__ value_cache,  // [num_blocks, kv_num_heads, block_size,
-                                  // head_size // 2]
-    T* __restrict__ qkv_out,
-    const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
-    const int* __restrict__ seq_lens,          // [bsz]
-    const int* __restrict__ seq_lens_encoder,  // [bsz]
-    const float* __restrict__ cos_emb,
-    const float* __restrict__ sin_emb,
-    const int max_seq_len,
-    const int max_blocks_per_seq,
-    const int num_heads,
-    const int head_size,
-    const int block_size,
-    const uint32_t elem_cnt,
-    const int kv_num_heads,
-    const bool rope_3d,
-    const float* q_norm_weight,
-    const float* k_norm_weight,
-    const float rms_norm_eps) {
-  using LoadT = AlignedVector<T, VecSize>;
-  using LoadBiasT = AlignedVector<T, VecSize>;
-  using LoadKVT = AlignedVector<T, VecSize>;
-  constexpr int HalfVecSize = VecSize / 2;
-  using LoadEmbT = AlignedVector<float, HalfVecSize>;
-  using LoadFloat = AlignedVector<float, VecSize>;
-  LoadT src_vec;
-  LoadBiasT out_vec;
-  LoadKVT cache_vec;
-  LoadEmbT cos_emb_vec;
-  LoadEmbT sin_emb_vec;
-  LoadFloat tmp_vec;
-  LoadFloat q_norm_vec, k_norm_vec;
-
-  int64_t global_warp_idx = blockDim.y * blockIdx.x + threadIdx.y;
-  int64_t all_warp_num = gridDim.x * blockDim.y;
-  int64_t all_head_dim = elem_cnt / head_size;
-
-  const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * head_size;
-  const int half_head_size = head_size / 2;
-  for (int gloabl_hi = global_warp_idx; gloabl_hi < all_head_dim; gloabl_hi += all_warp_num) {
-    int64_t linear_index = gloabl_hi * head_size + threadIdx.x * VecSize;
-    const int ori_bi = linear_index / hidden_size;
-    const int bias = linear_index % hidden_size;
-    const int hi = bias / head_size;  // q + k + v
-    const int h_bias = bias % head_size;
-    const int start_token_idx = cu_seqlens_q[ori_bi];
-    if (seq_lens_encoder[ori_bi] > 0) return;
-    const int write_seq_id = seq_lens[ori_bi];
-    if (write_seq_id == 0) continue;
-
-    const int* block_table_now = nullptr;
-
-    block_table_now = block_tables + ori_bi * max_blocks_per_seq;
-    const int block_idx = block_table_now[write_seq_id / block_size];
-    const int block_offset = write_seq_id % block_size;
-    const uint32_t ori_idx =
-        start_token_idx * hidden_size + hi * head_size + h_bias;
-
-    const int bias_idx = hi * head_size + h_bias;
-    Load<T, VecSize>(&quant_qkv[ori_idx], &src_vec);
-    if (hi < num_heads + kv_num_heads) {
-      // q k rope
-      const uint32_t emb_idx = write_seq_id * half_head_size + h_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + ori_bi * max_seq_len * head_size : emb_idx;
-      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
-    }
-    float thread_m2 = 0.0f;
-    float warp_m2 = 0.0f;
-
-#pragma unroll
-    for (int i = 0; i < HalfVecSize; i++) {
-      // dequant + add_bias + rope
-      float input_left = static_cast<float>(src_vec[2 * i]);
-      float input_right = static_cast<float>(src_vec[2 * i + 1]);
-
-      if (hi < num_heads + kv_num_heads) {
-        const float cos_tmp = cos_emb_vec[i];
-        const float sin_tmp = sin_emb_vec[i];
-        float tmp1 = input_left * cos_tmp - input_right * sin_tmp;
-        float tmp2 = input_right * cos_tmp + input_left * sin_tmp;
-        thread_m2 += tmp1 * tmp1 + tmp2 * tmp2;
-        tmp_vec[2 * i] = tmp1;
-        tmp_vec[2 * i + 1] = tmp2;
-      } else {
-        out_vec[2 * i] = src_vec[2 * i];
-        out_vec[2 * i + 1] = src_vec[2 * i + 1];
-      }
-    }
-    if (hi < (num_heads + kv_num_heads)) { // q k
-      WelfordWarpAllReduce<float, 32>(thread_m2, &warp_m2);
-      float row_variance =
-          max(warp_m2 / head_size, 0.0f);
-      float row_inv_var = Rsqrt(row_variance + rms_norm_eps);
-      if (hi < num_heads) { // q
-        Load<float, VecSize>(&q_norm_weight[threadIdx.x * VecSize], &q_norm_vec);
-        #pragma unroll
-        for (int i = 0; i < VecSize; i++) {
-          out_vec[i] = static_cast<T>(tmp_vec[i] * row_inv_var * q_norm_vec[i]);
-        }
-      } else { // k
-        Load<float, VecSize>(&k_norm_weight[threadIdx.x * VecSize], &k_norm_vec);
-        #pragma unroll
-        for (int i = 0; i < VecSize; i++) {
-          out_vec[i] = static_cast<T>(tmp_vec[i] * row_inv_var * k_norm_vec[i]);
-        }
-      }
-    }
-    if (hi < num_heads) {
-      // write q
-      Store<T, VecSize>(out_vec, &qkv_out[ori_idx]);
-    } else {
-      // quant + write k/v
-      const uint32_t kv_head_idx = (hi - num_heads) % kv_num_heads;
-      const uint32_t tgt_idx =
-          block_idx * kv_num_heads * block_size * head_size +
-          kv_head_idx * block_size * head_size + block_offset * head_size +
-          h_bias;
-      if (hi < num_heads + kv_num_heads) {
-        Store<T, VecSize>(out_vec, &key_cache[tgt_idx]);
-      } else {
-        Store<T, VecSize>(out_vec, &value_cache[tgt_idx]);
-      }
-    }
-
-  }
-}
-
 template <typename T, int VecSize = 1>
 __global__ void append_decode_cache_T_rope_kernel(
     const T* __restrict__ quant_qkv,  // [bsz, num_heads + 2 * kv_num_heads,
@@ -164,8 +28,8 @@ __global__ void append_decode_cache_T_rope_kernel(
                                   // head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -201,7 +65,7 @@ __global__ void append_decode_cache_T_rope_kernel(
     const int bias = linear_index % hidden_size;
     const int hi = bias / head_size;  // q + k + v
     const int h_bias = bias % head_size;
-    const int start_token_idx = cu_seqlens_q[ori_bi];
+    const int start_token_idx = ori_bi * max_seq_len - cum_offsets[ori_bi];
     if (seq_lens_encoder[ori_bi] > 0) return;
     const int write_seq_id = seq_lens[ori_bi];
     if (write_seq_id == 0) continue;
@@ -270,8 +134,8 @@ __global__ void append_decode_cache_T_rope_kernel(
                                   // head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -313,7 +177,7 @@ __global__ void append_decode_cache_T_rope_kernel(
     const int bias = linear_index % hidden_size;
     const int hi = bias / head_size;  // q + k + v
     const int h_bias = bias % head_size;
-    const int start_token_idx = cu_seqlens_q[ori_bi];
+    const int start_token_idx = ori_bi * max_seq_len - cum_offsets[ori_bi];
     if (seq_lens_encoder[ori_bi] > 0) return;
     const int write_seq_id = seq_lens[ori_bi];
     if (write_seq_id == 0) continue;
@@ -335,9 +199,8 @@ __global__ void append_decode_cache_T_rope_kernel(
     if (hi < num_heads + kv_num_heads) {
       // q k rope
       const uint32_t emb_idx = write_seq_id * half_head_size + h_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + ori_bi * max_seq_len * head_size : emb_idx;
-      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
     }
 #pragma unroll
     for (int i = 0; i < HalfVecSize; i++) {
@@ -391,8 +254,8 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
                                   // head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -403,8 +266,7 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
     const int head_size,
     const int block_size,
     const uint32_t elem_cnt,
-    const int kv_num_heads,
-    const bool rope_3d) {
+    const int kv_num_heads) {
   using LoadT = AlignedVector<T, VecSize>;
   using LoadBiasT = AlignedVector<T, VecSize>;
   using LoadKVT = AlignedVector<T, VecSize>;
@@ -431,7 +293,7 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
     const int bias = linear_index % half_hidden_size;
     const int hi = bias / half_head_size;  // q + k + v
     const int h_bias = bias % half_head_size;
-    const int start_token_idx = cu_seqlens_q[ori_bi];
+    const int start_token_idx = ori_bi * max_seq_len - cum_offsets[ori_bi];
     if (seq_lens_encoder[ori_bi] > 0) return;
     const int write_seq_id = seq_lens[ori_bi];
     if (write_seq_id == 0) continue;
@@ -451,9 +313,8 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
     if (hi < num_heads + kv_num_heads) {
       // q k rope
       const uint32_t emb_idx = write_seq_id * head_size + h_bias;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + ori_bi * max_seq_len * head_size * 2 : emb_idx;
-      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
     }
 #pragma unroll
     for (int i = 0; i < VecSize; i++) {
@@ -505,8 +366,8 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
                                   // head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -521,8 +382,7 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
     const int head_size,
     const int block_size,
     const uint32_t elem_cnt,
-    const int kv_num_heads,
-    const bool rope_3d) {
+    const int kv_num_heads) {
   using LoadT = AlignedVector<int, VecSize>;
   using LoadBiasT = AlignedVector<T, VecSize>;
   using LoadOutScaleT = AlignedVector<float, VecSize>;
@@ -549,7 +409,7 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
     const int bias = linear_index % half_hidden_size;
     const int hi = bias / half_head_size;  // q + k + v
     const int h_bias = bias % half_head_size;
-    const int start_token_idx = cu_seqlens_q[ori_bi];
+    const int start_token_idx = ori_bi * max_seq_len - cum_offsets[ori_bi];
     if (seq_lens_encoder[ori_bi] > 0) return;
     const int write_seq_id = seq_lens[ori_bi];
     if (write_seq_id == 0) continue;
@@ -579,9 +439,8 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
     if (hi < num_heads + kv_num_heads) {
       // q k rope
       const uint32_t emb_idx = write_seq_id * head_size + h_bias;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + ori_bi * max_seq_len * head_size * 2 : emb_idx;
-      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
     }
 #pragma unroll
     for (int i = 0; i < VecSize; i++) {
@@ -629,294 +488,6 @@ __global__ void append_decode_cache_T_neox_rope_kernel(
   }
 }
 
-template <typename T, int VecSize = 4, int RoundType = 0, int HeadDim = 128, bool is_scale_channel_wise=false, bool IsFP8=true>
-__global__ void append_decode_cache_int8_rope_qk_norm_kernel(
-    const T* __restrict__ quant_qkv,    // [bsz, num_heads + 2 * kv_num_heads,
-                                        // head_size]
-    uint8_t* __restrict__ key_cache,    // [num_blocks, kv_num_heads,
-                                        // block_size, head_size // 2]
-    uint8_t* __restrict__ value_cache,  // [num_blocks, kv_num_heads,
-                                        // block_size, head_size // 2]
-    T* __restrict__ qkv_out,
-    const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
-    const int* __restrict__ seq_lens,          // [bsz]
-    const int* __restrict__ seq_lens_encoder,  // [bsz]
-    const float* __restrict__ cos_emb,
-    const float* __restrict__ sin_emb,
-    T* __restrict__ cache_k_scale,
-    T* __restrict__ cache_v_scale,
-    const float* q_norm_weight,
-    const float* k_norm_weight,
-    const int max_seq_len,
-    const int max_blocks_per_seq,
-    const int num_heads,
-    const int block_size,
-    const float max_bound,
-    const float min_bound,
-    const int kv_num_heads,
-    const bool rope_3d,
-    const float rms_norm_eps) {
-  static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
-  static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
-  constexpr int NUM_WARPS = 4;
-  const int tid = threadIdx.x;
-  const int wid = tid / 32;
-  const int lane_id = tid % 32;
-  const int bid = blockIdx.x, head_idx = blockIdx.y * NUM_WARPS + wid;
-  int q_head_idx, k_head_idx, v_idx;
-  const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * HeadDim;
-  constexpr int half_head_size = HeadDim / 2;
-  const int start_token_idx = cu_seqlens_q[bid];
-  if (seq_lens_encoder[bid] > 0) return;
-  const int write_seq_id = seq_lens[bid];
-  if (write_seq_id == 0) return;
-  const int* block_table_now = nullptr;
-
-  block_table_now = block_tables + bid * max_blocks_per_seq;
-  const int block_idx = __ldg(&block_table_now[write_seq_id / block_size]);
-  const int block_offset = write_seq_id % block_size;
-
-  int cache_offset;
-  if (head_idx < num_heads) {
-    cache_offset = 0;
-  } else if (head_idx < num_heads + 2 * kv_num_heads) {
-    cache_offset = block_idx * kv_num_heads * block_size + (head_idx - num_heads) % kv_num_heads * block_size + block_offset;
-  }
-  T *cache_k_scale_now = cache_k_scale + cache_offset;
-  T *cache_v_scale_now = cache_v_scale + cache_offset;
-
-  float thread_m2 = 0.0f;
-  float warp_m2 = 0.0f;
-
-  if (head_idx < num_heads) {
-    // q
-    using LoadT = AlignedVector<T, VecSize>;
-    using LoadBiasT = AlignedVector<T, VecSize>;
-    using LoadOutScaleT = AlignedVector<float, VecSize>;
-    constexpr int HalfVecSize = VecSize / 2;
-    using LoadEmbT = AlignedVector<float, HalfVecSize>;
-
-    LoadT src_vec;
-    LoadBiasT out_vec;
-    LoadEmbT cos_emb_vec;
-    LoadEmbT sin_emb_vec;
-    const T* qkv_now = quant_qkv + start_token_idx * hidden_size;
-    T* qkv_out_now = qkv_out + start_token_idx * hidden_size;
-#pragma unroll
-    for (uint32_t head_bias = lane_id * VecSize; head_bias < HeadDim;
-         head_bias += 32 * VecSize) {
-      const int bias_idx = head_idx * HeadDim + head_bias;
-      Load<T, VecSize>(&qkv_now[bias_idx], &src_vec);
-      // q rope
-      const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      const uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
-#pragma unroll
-      for (int i = 0; i < HalfVecSize; i++) {
-        // dequant + add_bias + rope
-        float input_left = static_cast<float>(src_vec[2 * i]);
-        float input_right = static_cast<float>(src_vec[2 * i + 1]);
-
-        const float cos_tmp = cos_emb_vec[i];
-        const float sin_tmp = sin_emb_vec[i];
-        float tmp1 = input_left * cos_tmp - input_right * sin_tmp;
-        float tmp2 = input_right * cos_tmp + input_left * sin_tmp;
-        thread_m2 += tmp1 * tmp1 + tmp2 * tmp2;
-        out_vec[2 * i] =
-            static_cast<T>(tmp1);
-        out_vec[2 * i + 1] =
-            static_cast<T>(tmp2);
-      }
-      // qk norm
-      if (q_norm_weight) {
-        WelfordWarpAllReduce<float, 32>(thread_m2, &warp_m2);
-        float row_variance =
-            max(warp_m2 / HeadDim, 0.0f);
-        float row_inv_var = Rsqrt(row_variance + rms_norm_eps);
-        LoadOutScaleT q_norm_vec;
-        Load<float, VecSize>(&q_norm_weight[lane_id * VecSize], &q_norm_vec);
-        #pragma unroll
-        for (int i = 0; i < VecSize; i++) {
-          out_vec[i] = static_cast<T>(static_cast<float>(out_vec[i]) * row_inv_var * q_norm_vec[i]);
-        }
-      }
-      Store<T, VecSize>(out_vec, &qkv_out_now[bias_idx]);
-    }
-  } else if (head_idx < num_heads + 2 * kv_num_heads) {
-    // k
-    constexpr int KV_VEC_SIZE = 16 / sizeof(uint8_t);  // 16
-    using LoadPadKVT = AlignedVector<uint8_t, KV_VEC_SIZE>;
-    const uint32_t kv_head_idx = (head_idx - num_heads) % kv_num_heads;
-    if (block_offset == 0) {
-      // pad zero for this kv_head_idx for this block
-      LoadPadKVT pad_cache_vec;
-      *(reinterpret_cast<uint4*>(pad_cache_vec.val)) = make_uint4(0, 0, 0, 0);
-      if (head_idx < num_heads + kv_num_heads) {
-        constexpr int num_vecs_per_head_dim = HeadDim / KV_VEC_SIZE;
-        constexpr int num_token_each_time = 32 / num_vecs_per_head_dim;
-        const uint32_t tgt_idx =
-            (block_idx * kv_num_heads + kv_head_idx) * block_size * HeadDim +
-            lane_id % num_vecs_per_head_dim * KV_VEC_SIZE;
-        for (int block_i = lane_id / num_vecs_per_head_dim;
-             block_i < block_size;
-             block_i += num_token_each_time) {
-          Store<uint8_t, KV_VEC_SIZE>(pad_cache_vec,
-                                      &key_cache[tgt_idx + block_i * HeadDim]);
-        }
-      } else {
-        const int num_vecs_per_head_dim = block_size / KV_VEC_SIZE;
-        const int num_token_each_time = 32 / num_vecs_per_head_dim;
-        const uint32_t tgt_idx =
-            (block_idx * kv_num_heads + kv_head_idx) * HeadDim * block_size +
-            lane_id % num_vecs_per_head_dim * KV_VEC_SIZE;
-        for (int block_i = lane_id / num_vecs_per_head_dim; block_i < HeadDim;
-             block_i += num_token_each_time) {
-          Store<uint8_t, KV_VEC_SIZE>(
-              pad_cache_vec, &value_cache[tgt_idx + block_i * block_size]);
-        }
-      }
-      __syncwarp();
-    }
-
-    constexpr int K_VEC_SIZE = 4;
-    constexpr int HALF_K_VEC_SIZE = 2;
-    using LoadKVResT = AlignedVector<uint8_t, K_VEC_SIZE>;
-    using LoadKVT = AlignedVector<uint8_t, HALF_K_VEC_SIZE>;
-    using LoadT = AlignedVector<T, HALF_K_VEC_SIZE>;
-    using LoadBiasT = AlignedVector<T, HALF_K_VEC_SIZE>;
-    using LoadOutScaleT = AlignedVector<float, HALF_K_VEC_SIZE>;
-    using LoadEmbT = AlignedVector<float, 1>;
-    LoadKVResT cache_vec;
-    LoadT src_vec1, src_vec2;
-    LoadBiasT out_vec1, out_vec2;
-    LoadEmbT cos_emb_vec1, cos_emb_vec2;
-    LoadEmbT sin_emb_vec1, sin_emb_vec2;
-
-    const T* qkv_now = quant_qkv + start_token_idx * hidden_size;
-    const int head_bias = lane_id / 4 * 16 + lane_id % 4 * 2;
-    const int bias_idx = head_idx * HeadDim + head_bias;
-    Load<T, HALF_K_VEC_SIZE>(&qkv_now[bias_idx], &src_vec1);
-    Load<T, HALF_K_VEC_SIZE>(&qkv_now[bias_idx + 8], &src_vec2);
-    T scale = T(1.0f);
-    const int k_head_idx = head_idx - num_heads;
-    const int v_head_idx = head_idx - num_heads - kv_num_heads;
-    if (head_idx < num_heads + kv_num_heads) {
-      const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      const uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-      Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
-      Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-      Load<float, 1>(&sin_emb[new_emb_idx + 4], &sin_emb_vec2);
-    }
-
-    float input_left = static_cast<float>(src_vec1[0]);
-    float input_right = static_cast<float>(src_vec1[1]);
-    if (head_idx < num_heads + kv_num_heads) {
-      float cos_tmp = cos_emb_vec1[0];
-      float sin_tmp = sin_emb_vec1[0];
-      float tmp1 = input_left * cos_tmp - input_right * sin_tmp;
-      float tmp2 = input_right * cos_tmp + input_left * sin_tmp;
-      thread_m2 += tmp1 * tmp1 + tmp2 * tmp2;
-      out_vec1[0] =
-          static_cast<T>(tmp1);
-      out_vec1[1] =
-          static_cast<T>(tmp2);
-    } else {
-      out_vec1[0] = src_vec1[0];
-      out_vec1[1] = src_vec1[1];
-    }
-
-    // rope
-    input_left = static_cast<float>(src_vec2[0]);
-    input_right = static_cast<float>(src_vec2[1]);
-    if (head_idx < num_heads + kv_num_heads) {
-      float cos_tmp = cos_emb_vec2[0];
-      float sin_tmp = sin_emb_vec2[0];
-      float tmp1 = input_left * cos_tmp - input_right * sin_tmp;
-      float tmp2 = input_right * cos_tmp + input_left * sin_tmp;
-      thread_m2 += tmp1 * tmp1 + tmp2 * tmp2;
-      out_vec2[0] = static_cast<T>(tmp1);
-      out_vec2[1] = static_cast<T>(tmp2);
-    } else {
-      out_vec2[0] = src_vec2[0];
-      out_vec2[1] = src_vec2[1];
-    }
-    if (k_norm_weight) {
-      if (head_idx < num_heads + kv_num_heads) {
-        LoadOutScaleT k_norm_vec1, k_norm_vec2;
-        Load<float, HALF_K_VEC_SIZE>(&k_norm_weight[head_bias], &k_norm_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&k_norm_weight[head_bias + 8], &k_norm_vec2);
-        // qk norm
-        WelfordWarpAllReduce<float, 32>(thread_m2, &warp_m2);
-        float row_variance =
-            max(warp_m2 / HeadDim, 0.0f);
-        float row_inv_var = Rsqrt(row_variance + rms_norm_eps);
-
-        for (int i = 0; i < HALF_K_VEC_SIZE; i++) {
-          out_vec1[i] = static_cast<T>(static_cast<float>(out_vec1[i]) * row_inv_var * k_norm_vec1[i]);
-          out_vec2[i] = static_cast<T>(static_cast<float>(out_vec2[i]) * row_inv_var * k_norm_vec2[i]);
-        }
-      }
-    }
-    // reduce max, 1 head per warp
-    T local_max = -INFINITY;
-#pragma unroll
-    for (int i = 0; i < HALF_K_VEC_SIZE; i++) {
-      local_max = __hmax(local_max, __habs(out_vec1[i]));
-      local_max = __hmax(local_max, __habs(out_vec2[i]));
-    }
-#pragma unroll
-    for (int m_offset = 16; m_offset > 0; m_offset /= 2) {
-      local_max = __hmax(local_max, __shfl_xor_sync(0xffffffff, local_max, m_offset));
-    }
-
-    scale = __hdiv(448, local_max);
-
-    if (lane_id == 0) {
-      if (head_idx < num_heads + kv_num_heads) {
-        cache_k_scale_now[0] = __hdiv(1, scale);
-      } else {
-        cache_v_scale_now[0] = __hdiv(1, scale);
-      }
-    }
-
-#pragma unroll
-    for (uint32_t i = 0; i < HALF_K_VEC_SIZE; i++) {
-      cache_vec[i] = QuantToC8<T,true, IsFP8, RoundType>(scale, out_vec1[i], max_bound, min_bound);
-      cache_vec[i + HALF_K_VEC_SIZE] = QuantToC8<T,true, IsFP8, RoundType>(scale, out_vec2[i], max_bound, min_bound);
-    }
-    if (head_idx < num_heads + kv_num_heads) {
-      const int start_block_16 =
-          block_offset / 16 * 16 + block_offset % 8 + lane_id / 4 % 2 * 8;
-      const uint32_t tgt_cache_idx =
-          block_idx * kv_num_heads * block_size * HeadDim +
-          kv_head_idx * block_size * HeadDim + start_block_16 * HeadDim +
-          lane_id / 4 / 2 * 32 + (block_offset % 16) / 8 * 16 + lane_id % 4 * 4;
-      Store<uint8_t, K_VEC_SIZE>(cache_vec, &key_cache[tgt_cache_idx]);
-    } else {
-      const uint32_t base_tgt_cache_idx =
-          block_idx * kv_num_heads * HeadDim * block_size +
-          kv_head_idx * HeadDim * block_size +
-          (lane_id / 4 * 16 + lane_id % 4 * 2) * block_size +
-          block_offset / 16 % 2 * 8 * block_size + block_offset / 16 / 2 * 32;
-      const uint32_t tgt_cache_idx1 = base_tgt_cache_idx +
-                                      block_offset % 8 / 2 * 4     // per 4
-                                      + block_offset % 16 / 8 * 2  // per 2
-                                      + block_offset % 2;          // per 1
-      const uint32_t tgt_cache_idx2 = tgt_cache_idx1 + block_size;
-      const uint32_t tgt_cache_idx3 = tgt_cache_idx1 + 16;
-      const uint32_t tgt_cache_idx4 = tgt_cache_idx3 + block_size;
-      value_cache[tgt_cache_idx1] = cache_vec[0];
-      value_cache[tgt_cache_idx2] = cache_vec[1];
-      value_cache[tgt_cache_idx3] = cache_vec[2];
-      value_cache[tgt_cache_idx4] = cache_vec[3];
-    }
-  }
-}
-
 template <typename T, int VecSize = 4, int RoundType = 0, int HeadDim = 128, bool is_scale_channel_wise=false, bool IsFP8=false>
 __global__ void append_decode_cache_int8_rope_kernel(
     const T* __restrict__ quant_qkv,    // [bsz, num_heads + 2 * kv_num_heads,
@@ -927,8 +498,8 @@ __global__ void append_decode_cache_int8_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -941,8 +512,7 @@ __global__ void append_decode_cache_int8_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads,
-    const bool rope_3d) {
+    const int kv_num_heads) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -953,7 +523,7 @@ __global__ void append_decode_cache_int8_rope_kernel(
   int q_head_idx, k_head_idx, v_idx;
   const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * HeadDim;
   constexpr int half_head_size = HeadDim / 2;
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - __ldg(&cum_offsets[bid]);
   if (seq_lens_encoder[bid] > 0) return;
   const int write_seq_id = seq_lens[bid];
   if (write_seq_id == 0) return;
@@ -985,9 +555,8 @@ __global__ void append_decode_cache_int8_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < HalfVecSize; i++) {
         // dequant + add_bias + rope
@@ -1064,11 +633,10 @@ __global__ void append_decode_cache_int8_rope_kernel(
     const T *cache_v_scale_cur = cache_v_scale + v_head_idx * HeadDim + head_bias;
     if (head_idx < num_heads + kv_num_heads) {
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-      Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
-      Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-      Load<float, 1>(&sin_emb[new_emb_idx + 4], &sin_emb_vec2);
+      Load<float, 1>(&cos_emb[emb_idx], &cos_emb_vec1);
+      Load<float, 1>(&cos_emb[emb_idx + 4], &cos_emb_vec2);
+      Load<float, 1>(&sin_emb[emb_idx], &sin_emb_vec1);
+      Load<float, 1>(&sin_emb[emb_idx + 4], &sin_emb_vec2);
       if constexpr (!is_scale_channel_wise) {
         scale = __ldg(&cache_k_scale[kv_head_idx]);
       }
@@ -1177,8 +745,8 @@ __global__ void append_decode_cache_int8_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -1195,8 +763,7 @@ __global__ void append_decode_cache_int8_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads,
-    const bool rope_3d) {
+    const int kv_num_heads) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -1208,7 +775,7 @@ __global__ void append_decode_cache_int8_rope_kernel(
   int q_head_idx, k_head_idx, v_idx;
   const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * HeadDim;
   constexpr int half_head_size = HeadDim / 2;
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - __ldg(&cum_offsets[bid]);
   if (seq_lens_encoder[bid] > 0) return;
   const int write_seq_id = seq_lens[bid];
   if (write_seq_id == 0) return;
@@ -1246,10 +813,9 @@ __global__ void append_decode_cache_int8_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
 
-      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
 
 #pragma unroll
       for (int i = 0; i < HalfVecSize; i++) {
@@ -1342,11 +908,10 @@ __global__ void append_decode_cache_int8_rope_kernel(
     const T *cache_v_scale_cur = cache_v_scales + v_head_idx * HeadDim + head_bias;
     if (head_idx < num_heads + kv_num_heads) {
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-      Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
-      Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-      Load<float, 1>(&sin_emb[new_emb_idx + 4], &sin_emb_vec2);
+      Load<float, 1>(&cos_emb[emb_idx], &cos_emb_vec1);
+      Load<float, 1>(&cos_emb[emb_idx + 4], &cos_emb_vec2);
+      Load<float, 1>(&sin_emb[emb_idx], &sin_emb_vec1);
+      Load<float, 1>(&sin_emb[emb_idx + 4], &sin_emb_vec2);
       if constexpr (!is_scale_channel_wise) {
         scale = __ldg(&cache_k_scales[kv_head_idx]);
       }
@@ -1482,8 +1047,8 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -1496,8 +1061,7 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads,
-    const bool rope_3d) {
+    const int kv_num_heads) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -1509,7 +1073,7 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
   int q_head_idx, k_head_idx, v_idx;
   const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * HeadDim;
   constexpr int half_head_size = HeadDim / 2;
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - __ldg(&cum_offsets[bid]);
   if (seq_lens_encoder[bid] > 0) return;
   const int write_seq_id = seq_lens[bid];
   if (write_seq_id == 0) return;
@@ -1545,9 +1109,8 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
-      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
 
 #pragma unroll
       for (int i = 0; i < VecSize; i++) {
@@ -1628,11 +1191,10 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
 
         T scale;
         const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-        uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx + 8], &cos_emb_vec2);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx + 8], &sin_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx], &cos_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx + 8], &cos_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx], &sin_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx + 8], &sin_emb_vec2);
         scale = __ldg(&cache_k_scales[kv_head_idx]);
 #pragma unroll
         for (int i = 0; i < HALF_K_VEC_SIZE; i++) {
@@ -1784,8 +1346,8 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -1802,8 +1364,7 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads,
-    const bool rope_3d) {
+    const int kv_num_heads) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -1816,7 +1377,7 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
 
   const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * HeadDim;
   constexpr int half_head_size = HeadDim / 2;
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - __ldg(&cum_offsets[bid]);
   if (seq_lens_encoder[bid] > 0) return;
   const int write_seq_id = seq_lens[bid];
   if (write_seq_id == 0) return;
@@ -1863,10 +1424,8 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
-      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
-
+      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
 
 #pragma unroll
       for (int i = 0; i < VecSize; i++) {
@@ -1974,11 +1533,10 @@ __global__ void append_decode_cache_int8_neox_rope_kernel(
 
         T scale;
         const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-        uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx + 8], &cos_emb_vec2);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx + 8], &sin_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx], &cos_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx + 8], &cos_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx], &sin_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx + 8], &sin_emb_vec2);
         scale = __ldg(&cache_k_scales[kv_head_idx]);
 #pragma unroll
         for (int i = 0; i < HALF_K_VEC_SIZE; i++) {
@@ -2181,8 +1739,8 @@ __global__ void append_decode_cache_int4_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -2197,8 +1755,7 @@ __global__ void append_decode_cache_int4_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads,
-    const bool rope_3d) {
+    const int kv_num_heads) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -2209,7 +1766,7 @@ __global__ void append_decode_cache_int4_rope_kernel(
   const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * HeadDim;
   constexpr int half_head_size = HeadDim / 2;
   const int half_block_size = block_size / 2;
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - __ldg(&cum_offsets[bid]);
   if (seq_lens_encoder[bid] > 0) return;
   const int write_seq_id = seq_lens[bid];
   if (write_seq_id == 0) return;
@@ -2242,9 +1799,8 @@ __global__ void append_decode_cache_int4_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < HalfVecSize; i++) {
         // dequant + add_bias + rope
@@ -2318,11 +1874,10 @@ __global__ void append_decode_cache_int4_rope_kernel(
     Load<T, HALF_K_VEC_SIZE>(&qkv_now[bias_idx + 8], &src_vec2);
     if (head_idx < num_heads + kv_num_heads) {
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-      Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
-      Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-      Load<float, 1>(&sin_emb[new_emb_idx + 4], &sin_emb_vec2);
+      Load<float, 1>(&cos_emb[emb_idx], &cos_emb_vec1);
+      Load<float, 1>(&cos_emb[emb_idx + 4], &cos_emb_vec2);
+      Load<float, 1>(&sin_emb[emb_idx], &sin_emb_vec1);
+      Load<float, 1>(&sin_emb[emb_idx + 4], &sin_emb_vec2);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[cache_idx], &scale_vec1);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[cache_idx + 8], &scale_vec2);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_zero_points[cache_idx], &zp_vec1);
@@ -2479,8 +2034,8 @@ __global__ void append_decode_cache_int4_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -2499,8 +2054,7 @@ __global__ void append_decode_cache_int4_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads,
-    const bool rope_3d) {
+    const int kv_num_heads) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -2512,7 +2066,7 @@ __global__ void append_decode_cache_int4_rope_kernel(
   const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * HeadDim;
   constexpr int half_head_size = HeadDim / 2;
   const int half_block_size = block_size / 2;
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - __ldg(&cum_offsets[bid]);
   if (seq_lens_encoder[bid] > 0) return;
   const int write_seq_id = seq_lens[bid];
   if (write_seq_id == 0) return;
@@ -2549,9 +2103,8 @@ __global__ void append_decode_cache_int4_rope_kernel(
       Load<float, VecSize>(&qkv_out_scales[bias_idx], &out_scale_vec);
       // q rope
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < HalfVecSize; i++) {
         // dequant + add_bias + rope
@@ -2638,11 +2191,10 @@ __global__ void append_decode_cache_int4_rope_kernel(
                                  &out_scale_vec2);
     if (head_idx < num_heads + kv_num_heads) {
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-      Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
-      Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-      Load<float, 1>(&sin_emb[new_emb_idx + 4], &sin_emb_vec2);
+      Load<float, 1>(&cos_emb[emb_idx], &cos_emb_vec1);
+      Load<float, 1>(&cos_emb[emb_idx + 4], &cos_emb_vec2);
+      Load<float, 1>(&sin_emb[emb_idx], &sin_emb_vec1);
+      Load<float, 1>(&sin_emb[emb_idx + 4], &sin_emb_vec2);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[cache_idx], &scale_vec1);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[cache_idx + 8], &scale_vec2);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_zero_points[cache_idx], &zp_vec1);
@@ -2810,8 +2362,8 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -2826,8 +2378,7 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads,
-    const bool rope_3d) {
+    const int kv_num_heads) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -2838,7 +2389,7 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
   const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * HeadDim;
   constexpr int half_head_size = HeadDim / 2;
   const int half_block_size = block_size / 2;
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - __ldg(&cum_offsets[bid]);
   if (seq_lens_encoder[bid] > 0) return;
   const int write_seq_id = seq_lens[bid];
   if (write_seq_id == 0) return;
@@ -2874,9 +2425,8 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
-      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < VecSize; i++) {
         // dequant + add_bias + rope
@@ -2957,11 +2507,10 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
         Load<T, HALF_K_VEC_SIZE>(&qkv_now[right_bias_idx], &right_src_vec1);
         Load<T, HALF_K_VEC_SIZE>(&qkv_now[right_bias_idx + 8], &right_src_vec2);
         const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-        uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx + 8], &cos_emb_vec2);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx + 8], &sin_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx], &cos_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx + 8], &cos_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx], &sin_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx + 8], &sin_emb_vec2);
         Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[left_cache_idx],
                                  &left_scale_vec1);
         Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[left_cache_idx + 8],
@@ -3183,8 +2732,8 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -3203,8 +2752,7 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int kv_num_heads,
-    const bool rope_3d) {
+    const int kv_num_heads) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -3216,7 +2764,7 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
   const int64_t hidden_size = (num_heads + 2 * kv_num_heads) * HeadDim;
   constexpr int half_head_size = HeadDim / 2;
   const int half_block_size = block_size / 2;
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - __ldg(&cum_offsets[bid]);
   if (seq_lens_encoder[bid] > 0) return;
   const int write_seq_id = seq_lens[bid];
   if (write_seq_id == 0) return;
@@ -3262,9 +2810,8 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
                            &right_out_scale_vec);
       // q rope
       const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
-      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < VecSize; i++) {
         // dequant + add_bias + rope
@@ -3373,11 +2920,10 @@ __global__ void append_decode_cache_int4_neox_rope_kernel(
                                      &right_out_scale_vec2);
 
         const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-        uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx + 8], &cos_emb_vec2);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx + 8], &sin_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx], &cos_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx + 8], &cos_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx], &sin_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx + 8], &sin_emb_vec2);
         Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[left_cache_idx],
                                  &left_scale_vec1);
         Load<T, HALF_K_VEC_SIZE>(&cache_k_scale[left_cache_idx + 8],

custom_ops/gpu_ops/append_attn/decoder_write_cache_with_rope_kernel.cu

@@ -15,77 +15,14 @@
 #include "decoder_write_cache_with_rope_kernel.h"
 #include "utils.cuh"
 
-template <typename T, typename QKV_TYPE>
-void append_decode_cache_rope_qk_norm(const QKV_TYPE* qkv,
-                              T* key_cache,
-                              T* value_cache,
-                              T* qkv_out,
-                              const int* block_tables,
-                              const int* batch_id_per_token,
-                              const int* cu_seqlens_q,
-                              const int* seq_lens,
-                              const int* seq_lens_encoder,
-                              const float* cos_emb,
-                              const float* sin_emb,
-                              const float* qkv_out_scales,
-                              const T* qkv_biases,
-                              const int max_seq_len,
-                              const int max_blocks_per_seq,
-                              const int num_heads,
-                              const int kv_num_heads,
-                              const int dim_head,
-                              const int block_size,
-                              const int bsz,
-                              const cudaStream_t& stream,
-                              const bool use_neox_style,
-                              const bool rope_3d,
-                              const float* q_norm_weight,
-                              const float* k_norm_weight,
-                              const float rms_norm_eps) {
-  const uint32_t elem_nums =
-      use_neox_style ? bsz * (num_heads + 2 * kv_num_heads) * dim_head / 2
-                     : bsz * (num_heads + 2 * kv_num_heads) * dim_head;
-  constexpr int HEAD_DIM = 128;
-
-  constexpr int PackSize = HEAD_DIM / kWarpSize;
-  const int pack_num = elem_nums / PackSize;
-  const int blocksize = 128;
-  int grid_size = 1;
-  GetNumBlocks<128>(pack_num, &grid_size);
-  dim3 block_dim(kWarpSize, blocksize / kWarpSize, 1);
-  append_decode_cache_T_rope_qk_norm_kernel<T, PackSize>
-      <<<grid_size, block_dim, 0, stream>>>(reinterpret_cast<const T*>(qkv),
-                                            key_cache,
-                                            value_cache,
-                                            qkv_out,
-                                            block_tables,
-                                            batch_id_per_token,
-                                            cu_seqlens_q,
-                                            seq_lens,
-                                            seq_lens_encoder,
-                                            cos_emb,
-                                            sin_emb,
-                                            max_seq_len,
-                                            max_blocks_per_seq,
-                                            num_heads,
-                                            dim_head,
-                                            block_size,
-                                            elem_nums,
-                                            kv_num_heads,
-                                            rope_3d,
-                                            q_norm_weight,
-                                            k_norm_weight,
-                                            rms_norm_eps);
-}
-
 template <typename T, typename QKV_TYPE>
 void append_decode_cache_rope(const QKV_TYPE* qkv,
                               T* key_cache,
                               T* value_cache,
                               T* qkv_out,
                               const int* block_tables,
-                              const int* batch_id_per_token,
-                              const int* cu_seqlens_q,
+                              const int* padding_offsets,
+                              const int* cum_offsets,
                               const int* seq_lens,
                               const int* seq_lens_encoder,
                               const float* cos_emb,
@@ -120,8 +57,8 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
-              cu_seqlens_q,
+              padding_offsets,
+              cum_offsets,
               seq_lens,
               seq_lens_encoder,
               cos_emb,
@@ -134,8 +71,7 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
               dim_head,
               block_size,
               elem_nums,
-              kv_num_heads,
-              rope_3d);
+              kv_num_heads);
     } else {
       append_decode_cache_T_neox_rope_kernel<T, PackSize>
           <<<grid_size, blocksize, 0, stream>>>(reinterpret_cast<const T*>(qkv),
@@ -143,8 +79,8 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
                                                 value_cache,
                                                 qkv_out,
                                                 block_tables,
-                                                batch_id_per_token,
-                                                cu_seqlens_q,
+                                                padding_offsets,
+                                                cum_offsets,
                                                 seq_lens,
                                                 seq_lens_encoder,
                                                 cos_emb,
@@ -155,8 +91,7 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
                                                 dim_head,
                                                 block_size,
                                                 elem_nums,
-                                                kv_num_heads,
-                                                rope_3d);
+                                                kv_num_heads);
     }
   } else {
     if (qkv_out_scales) {
@@ -167,8 +102,8 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
-              cu_seqlens_q,
+              padding_offsets,
+              cum_offsets,
               seq_lens,
               seq_lens_encoder,
               cos_emb,
@@ -190,8 +125,8 @@ void append_decode_cache_rope(const QKV_TYPE* qkv,
                                                 value_cache,
                                                 qkv_out,
                                                 block_tables,
-                                                batch_id_per_token,
-                                                cu_seqlens_q,
+                                                padding_offsets,
+                                                cum_offsets,
                                                 seq_lens,
                                                 seq_lens_encoder,
                                                 cos_emb,
@@ -214,8 +149,8 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
                                    uint8_t* value_cache,
                                    T* qkv_out,
                                    const int* block_tables,
-                                   const int* batch_id_per_token,
-                                   const int* cu_seqlens_q,
+                                   const int* padding_offsets,
+                                   const int* cum_offsets,
                                    const int* seq_lens,
                                    const int* seq_lens_encoder,
                                    const float* cos_emb,
@@ -247,8 +182,8 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
-              cu_seqlens_q,
+              padding_offsets,
+              cum_offsets,
               seq_lens,
               seq_lens_encoder,
               cos_emb,
@@ -263,8 +198,7 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               block_size,
               127.0f,
               -127.0f,
-              kv_num_heads,
-              rope_3d);
+              kv_num_heads);
     } else {
       append_decode_cache_int8_neox_rope_kernel<T, 4>
           <<<grids, num_warps * 32, 0, stream>>>(
@@ -273,8 +207,8 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
-              cu_seqlens_q,
+              padding_offsets,
+              cum_offsets,
               seq_lens,
               seq_lens_encoder,
               cos_emb,
@@ -287,8 +221,7 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               block_size,
               127.0f,
               -127.0f,
-              kv_num_heads,
-              rope_3d);
+              kv_num_heads);
     }
   } else {
     if (qkv_out_scales) {
@@ -299,8 +232,8 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
-              cu_seqlens_q,
+              padding_offsets,
+              cum_offsets,
               seq_lens,
               seq_lens_encoder,
               cos_emb,
@@ -315,8 +248,7 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               block_size,
               127.0f,
               -127.0f,
-              kv_num_heads,
-              rope_3d);
+              kv_num_heads);
     } else {
       append_decode_cache_int8_rope_kernel<T, 4, 0, 128, is_scale_channel_wise, IsFP8>
           <<<grids, num_warps * 32, 0, stream>>>(
@@ -325,8 +257,8 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
-              cu_seqlens_q,
+              padding_offsets,
+              cum_offsets,
               seq_lens,
               seq_lens_encoder,
               cos_emb,
@@ -339,8 +271,7 @@ void append_decode_cache_int8_rope(const QKV_TYPE* qkv,
               block_size,
               127.0f,
               -127.0f,
-              kv_num_heads,
-              rope_3d);
+              kv_num_heads);
     }
   }
 }
@@ -351,8 +282,8 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
                                    uint8_t* value_cache,
                                    T* qkv_out,
                                    const int* block_tables,
-                                   const int* batch_id_per_token,
-                                   const int* cu_seqlens_q,
+                                   const int* padding_offsets,
+                                   const int* cum_offsets,
                                    const int* seq_lens,
                                    const int* seq_lens_encoder,
                                    const float* cos_emb,
@@ -386,8 +317,8 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
-              cu_seqlens_q,
+              padding_offsets,
+              cum_offsets,
               seq_lens,
               seq_lens_encoder,
               cos_emb,
@@ -404,8 +335,7 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               block_size,
               7.0f,
               -8.0f,
-              kv_num_heads,
-              rope_3d);
+              kv_num_heads);
     } else {
       append_decode_cache_int4_neox_rope_kernel<T, 4>
           <<<grids, num_warps * 32, 0, stream>>>(
@@ -414,8 +344,8 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
-              cu_seqlens_q,
+              padding_offsets,
+              cum_offsets,
               seq_lens,
               seq_lens_encoder,
               cos_emb,
@@ -430,8 +360,7 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               block_size,
               7.0f,
               -8.0f,
-              kv_num_heads,
-              rope_3d);
+              kv_num_heads);
     }
   } else {
     if (qkv_out_scales) {
@@ -442,8 +371,8 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
-              cu_seqlens_q,
+              padding_offsets,
+              cum_offsets,
               seq_lens,
               seq_lens_encoder,
               cos_emb,
@@ -460,8 +389,7 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               block_size,
               7.0f,
               -8.0f,
-              kv_num_heads,
-              rope_3d);
+              kv_num_heads);
     } else {
       append_decode_cache_int4_rope_kernel<T, 4>
           <<<grids, num_warps * 32, 0, stream>>>(
@@ -470,8 +398,8 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               value_cache,
               qkv_out,
               block_tables,
-              batch_id_per_token,
-              cu_seqlens_q,
+              padding_offsets,
+              cum_offsets,
               seq_lens,
               seq_lens_encoder,
               cos_emb,
@@ -486,8 +414,7 @@ void append_decode_cache_int4_rope(const QKV_TYPE* qkv,
               block_size,
               7.0f,
               -8.0f,
-              kv_num_heads,
-              rope_3d);
+              kv_num_heads);
     }
   }
 }
@@ -497,8 +424,8 @@ void DecoderWriteCacheWithRoPEKernel(
     const paddle::Tensor& qkv,
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& qkv_out_scales,
@@ -514,10 +441,7 @@ void DecoderWriteCacheWithRoPEKernel(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps) {
+    paddle::Tensor* value_cache_out) {
   typedef cascade_attn_type_traits<T> traits_;
   typedef cascade_attn_type_traits<QKV_TYPE> qkt_nv_type_;
   typedef typename traits_::type DataType_;
@@ -540,17 +464,80 @@ void DecoderWriteCacheWithRoPEKernel(
             ? rotary_embs.get().data<float>() + max_seq_len * dim_head
             : rotary_embs.get().data<float>() + max_seq_len * dim_head / 2;
   }
-
-  if (q_norm_weight && k_norm_weight) {
-    if (cache_quant_type_str == "none") {
-      append_decode_cache_rope_qk_norm(
+  if (cache_quant_type_str == "none") {
+    append_decode_cache_rope(
+        reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+        reinterpret_cast<DataType_*>(key_cache_out->data<T>()),
+        reinterpret_cast<DataType_*>(value_cache_out->data<T>()),
+        reinterpret_cast<DataType_*>(qkv_out->data<T>()),
+        block_tables.data<int>(),
+        padding_offsets.data<int>(),
+        cum_offsets.data<int>(),
+        seq_lens.data<int>(),
+        seq_lens_encoder.data<int>(),
+        cos_emb,
+        sin_emb,
+        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+        qkv_biases ? reinterpret_cast<DataType_*>(
+                         const_cast<T*>(qkv_biases.get().data<T>()))
+                   : nullptr,
+        max_seq_len,
+        max_blocks_per_seq,
+        num_heads,
+        kv_num_heads,
+        dim_head,
+        block_size,
+        bsz,
+        stream,
+        use_neox_rotary_style,
+        rope_3d);
+  } else if (cache_quant_type_str == "cache_int8") {
+    bool is_scale_channel_wise = false;
+    if (cache_k_scale && cache_k_scale.get().dims()[0] == dim_head * kv_num_heads) {
+      is_scale_channel_wise = true;
+    }
+    if (is_scale_channel_wise) {
+      append_decode_cache_int8_rope<DataType_, QKV_TYPE, true>(
+        reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+        key_cache_out->data<uint8_t>(),
+        value_cache_out->data<uint8_t>(),
+        reinterpret_cast<DataType_*>(qkv_out->data<T>()),
+        block_tables.data<int>(),
+        padding_offsets.data<int>(),
+        cum_offsets.data<int>(),
+        seq_lens.data<int>(),
+        seq_lens_encoder.data<int>(),
+        cos_emb,
+        sin_emb,
+        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+        qkv_biases ? reinterpret_cast<DataType_*>(
+                         const_cast<T*>(qkv_biases.get().data<T>()))
+                   : nullptr,
+        cache_k_scale ? reinterpret_cast<DataType_*>(
+                            const_cast<T*>(cache_k_scale.get().data<T>()))
+                      : nullptr,
+        cache_v_scale ? reinterpret_cast<DataType_*>(
+                            const_cast<T*>(cache_v_scale.get().data<T>()))
+                      : nullptr,
+        max_seq_len,
+        max_blocks_per_seq,
+        num_heads,
+        kv_num_heads,
+        dim_head,
+        block_size,
+        bsz,
+        stream,
+        use_neox_rotary_style,
+        rope_3d);
+    } else {
+      append_decode_cache_int8_rope<DataType_, QKV_TYPE, false>(
           reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-          reinterpret_cast<DataType_*>(key_cache_out->data<T>()),
-          reinterpret_cast<DataType_*>(value_cache_out->data<T>()),
+          key_cache_out->data<uint8_t>(),
+          value_cache_out->data<uint8_t>(),
           reinterpret_cast<DataType_*>(qkv_out->data<T>()),
           block_tables.data<int>(),
-          batch_id_per_token.data<int>(),
-          cu_seqlens_q.data<int>(),
+          padding_offsets.data<int>(),
+          cum_offsets.data<int>(),
           seq_lens.data<int>(),
           seq_lens_encoder.data<int>(),
           cos_emb,
@@ -559,6 +546,12 @@ void DecoderWriteCacheWithRoPEKernel(
           qkv_biases ? reinterpret_cast<DataType_*>(
                           const_cast<T*>(qkv_biases.get().data<T>()))
                     : nullptr,
+          cache_k_scale ? reinterpret_cast<DataType_*>(
+                              const_cast<T*>(cache_k_scale.get().data<T>()))
+                        : nullptr,
+          cache_v_scale ? reinterpret_cast<DataType_*>(
+                              const_cast<T*>(cache_v_scale.get().data<T>()))
+                        : nullptr,
           max_seq_len,
           max_blocks_per_seq,
           num_heads,
@@ -568,253 +561,84 @@ void DecoderWriteCacheWithRoPEKernel(
           bsz,
           stream,
           use_neox_rotary_style,
-          rope_3d,
-          q_norm_weight ? q_norm_weight.get().data<float>() : nullptr,
-          k_norm_weight ? k_norm_weight.get().data<float>() : nullptr,
-          rms_norm_eps);
-    } else if (cache_quant_type_str == "block_wise_fp8") {
-      constexpr int num_warps = 4;
-      const int all_warps =
-          ((num_heads + 2 * kv_num_heads) + num_warps - 1) / num_warps * num_warps;
-      dim3 grids(bsz, all_warps / num_warps);
-      append_decode_cache_int8_rope_qk_norm_kernel<DataType_, 4, 0, 128, false, true>
-          <<<grids, num_warps * 32, 0, stream>>>(
-              reinterpret_cast<const DataType_*>(qkv_ptr),
-              key_cache_out->data<uint8_t>(),
-              value_cache_out->data<uint8_t>(),
-              reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-              block_tables.data<int>(),
-              batch_id_per_token.data<int>(),
-              cu_seqlens_q.data<int>(),
-              seq_lens.data<int>(),
-              seq_lens_encoder.data<int>(),
-              cos_emb,
-              sin_emb,
-              const_cast<DataType_*>(reinterpret_cast<const DataType_*>(cache_k_scale.get().data<T>())),
-              const_cast<DataType_*>(reinterpret_cast<const DataType_*>((cache_v_scale.get().data<T>()))),
-              q_norm_weight.get().data<float>(),
-              k_norm_weight.get().data<float>(),
-              max_seq_len,
-              max_blocks_per_seq,
-              num_heads,
-              block_size,
-              127.0f,
-              -127.0f,
-              kv_num_heads,
-              rope_3d,
-              rms_norm_eps);
-    } else {
-      PD_THROW(
-          "append_decode_cache_rope_qk_norm just supports cache_quant_type none/block_wise_fp8");
+          rope_3d);
     }
+  } else if (cache_quant_type_str == "cache_fp8")  {
+      append_decode_cache_int8_rope<DataType_, QKV_TYPE, false, true>(
+          reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+          key_cache_out->data<uint8_t>(),
+          value_cache_out->data<uint8_t>(),
+          reinterpret_cast<DataType_*>(qkv_out->data<T>()),
+          block_tables.data<int>(),
+          padding_offsets.data<int>(),
+          cum_offsets.data<int>(),
+          seq_lens.data<int>(),
+          seq_lens_encoder.data<int>(),
+          cos_emb,
+          sin_emb,
+          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+          qkv_biases ? reinterpret_cast<DataType_*>(
+                          const_cast<T*>(qkv_biases.get().data<T>()))
+                    : nullptr,
+          cache_k_scale ? reinterpret_cast<DataType_*>(
+                              const_cast<T*>(cache_k_scale.get().data<T>()))
+                        : nullptr,
+          cache_v_scale ? reinterpret_cast<DataType_*>(
+                              const_cast<T*>(cache_v_scale.get().data<T>()))
+                        : nullptr,
+          max_seq_len,
+          max_blocks_per_seq,
+          num_heads,
+          kv_num_heads,
+          dim_head,
+          block_size,
+          bsz,
+          stream,
+          use_neox_rotary_style,
+          rope_3d);
+  } else if (cache_quant_type_str == "cache_int4_zp") {
+    append_decode_cache_int4_rope(
+        reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+        key_cache_out->data<uint8_t>(),
+        value_cache_out->data<uint8_t>(),
+        reinterpret_cast<DataType_*>(const_cast<T*>(qkv_out->data<T>())),
+        block_tables.data<int>(),
+        padding_offsets.data<int>(),
+        cum_offsets.data<int>(),
+        seq_lens.data<int>(),
+        seq_lens_encoder.data<int>(),
+        cos_emb,
+        sin_emb,
+        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+        qkv_biases ? reinterpret_cast<DataType_*>(
+                         const_cast<T*>(qkv_biases.get().data<T>()))
+                   : nullptr,
+        cache_k_scale ? reinterpret_cast<DataType_*>(
+                            const_cast<T*>(cache_k_scale.get().data<T>()))
+                      : nullptr,
+        cache_v_scale ? reinterpret_cast<DataType_*>(
+                            const_cast<T*>(cache_v_scale.get().data<T>()))
+                      : nullptr,
+        cache_k_zp ? reinterpret_cast<DataType_*>(
+                         const_cast<T*>(cache_k_zp.get().data<T>()))
+                   : nullptr,
+        cache_v_zp ? reinterpret_cast<DataType_*>(
+                         const_cast<T*>(cache_v_zp.get().data<T>()))
+                   : nullptr,
+        max_seq_len,
+        max_blocks_per_seq,
+        num_heads,
+        kv_num_heads,
+        dim_head,
+        block_size,
+        bsz,
+        stream,
+        use_neox_rotary_style,
+        rope_3d);
   } else {
-    if (cache_quant_type_str == "none") {
-      append_decode_cache_rope(
-          reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-          reinterpret_cast<DataType_*>(key_cache_out->data<T>()),
-          reinterpret_cast<DataType_*>(value_cache_out->data<T>()),
-          reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-          block_tables.data<int>(),
-          batch_id_per_token.data<int>(),
-          cu_seqlens_q.data<int>(),
-          seq_lens.data<int>(),
-          seq_lens_encoder.data<int>(),
-          cos_emb,
-          sin_emb,
-          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-          qkv_biases ? reinterpret_cast<DataType_*>(
-                          const_cast<T*>(qkv_biases.get().data<T>()))
-                    : nullptr,
-          max_seq_len,
-          max_blocks_per_seq,
-          num_heads,
-          kv_num_heads,
-          dim_head,
-          block_size,
-          bsz,
-          stream,
-          use_neox_rotary_style,
-          rope_3d);
-    } else if (cache_quant_type_str == "cache_int8") {
-      bool is_scale_channel_wise = false;
-      if (cache_k_scale && cache_k_scale.get().dims()[0] == dim_head * kv_num_heads) {
-        is_scale_channel_wise = true;
-      }
-      if (is_scale_channel_wise) {
-        append_decode_cache_int8_rope<DataType_, QKV_TYPE, true>(
-          reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-          key_cache_out->data<uint8_t>(),
-          value_cache_out->data<uint8_t>(),
-          reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-          block_tables.data<int>(),
-          batch_id_per_token.data<int>(),
-          cu_seqlens_q.data<int>(),
-          seq_lens.data<int>(),
-          seq_lens_encoder.data<int>(),
-          cos_emb,
-          sin_emb,
-          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-          qkv_biases ? reinterpret_cast<DataType_*>(
-                          const_cast<T*>(qkv_biases.get().data<T>()))
-                    : nullptr,
-          cache_k_scale ? reinterpret_cast<DataType_*>(
-                              const_cast<T*>(cache_k_scale.get().data<T>()))
-                        : nullptr,
-          cache_v_scale ? reinterpret_cast<DataType_*>(
-                              const_cast<T*>(cache_v_scale.get().data<T>()))
-                        : nullptr,
-          max_seq_len,
-          max_blocks_per_seq,
-          num_heads,
-          kv_num_heads,
-          dim_head,
-          block_size,
-          bsz,
-          stream,
-          use_neox_rotary_style,
-          rope_3d);
-      } else {
-        append_decode_cache_int8_rope<DataType_, QKV_TYPE, false>(
-            reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-            key_cache_out->data<uint8_t>(),
-            value_cache_out->data<uint8_t>(),
-            reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-            block_tables.data<int>(),
-            batch_id_per_token.data<int>(),
-            cu_seqlens_q.data<int>(),
-            seq_lens.data<int>(),
-            seq_lens_encoder.data<int>(),
-            cos_emb,
-            sin_emb,
-            qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-            qkv_biases ? reinterpret_cast<DataType_*>(
-                            const_cast<T*>(qkv_biases.get().data<T>()))
-                      : nullptr,
-            cache_k_scale ? reinterpret_cast<DataType_*>(
-                                const_cast<T*>(cache_k_scale.get().data<T>()))
-                          : nullptr,
-            cache_v_scale ? reinterpret_cast<DataType_*>(
-                                const_cast<T*>(cache_v_scale.get().data<T>()))
-                          : nullptr,
-            max_seq_len,
-            max_blocks_per_seq,
-            num_heads,
-            kv_num_heads,
-            dim_head,
-            block_size,
-            bsz,
-            stream,
-            use_neox_rotary_style,
-            rope_3d);
-      }
-    } else if (cache_quant_type_str == "cache_fp8")  {
-        append_decode_cache_int8_rope<DataType_, QKV_TYPE, false, true>(
-            reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-            key_cache_out->data<uint8_t>(),
-            value_cache_out->data<uint8_t>(),
-            reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-            block_tables.data<int>(),
-            batch_id_per_token.data<int>(),
-            cu_seqlens_q.data<int>(),
-            seq_lens.data<int>(),
-            seq_lens_encoder.data<int>(),
-            cos_emb,
-            sin_emb,
-            qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-            qkv_biases ? reinterpret_cast<DataType_*>(
-                            const_cast<T*>(qkv_biases.get().data<T>()))
-                      : nullptr,
-            cache_k_scale ? reinterpret_cast<DataType_*>(
-                                const_cast<T*>(cache_k_scale.get().data<T>()))
-                          : nullptr,
-            cache_v_scale ? reinterpret_cast<DataType_*>(
-                                const_cast<T*>(cache_v_scale.get().data<T>()))
-                          : nullptr,
-            max_seq_len,
-            max_blocks_per_seq,
-            num_heads,
-            kv_num_heads,
-            dim_head,
-            block_size,
-            bsz,
-            stream,
-            use_neox_rotary_style,
-            rope_3d);
-    } else if (cache_quant_type_str == "block_wise_fp8") {
-      constexpr int num_warps = 4;
-      const int all_warps =
-          ((num_heads + 2 * kv_num_heads) + num_warps - 1) / num_warps * num_warps;
-      dim3 grids(bsz, all_warps / num_warps);
-      append_decode_cache_int8_rope_qk_norm_kernel<DataType_, 4, 0, 128, false, true>
-          <<<grids, num_warps * 32, 0, stream>>>(
-              reinterpret_cast<const DataType_*>(qkv_ptr),
-              key_cache_out->data<uint8_t>(),
-              value_cache_out->data<uint8_t>(),
-              reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-              block_tables.data<int>(),
-              batch_id_per_token.data<int>(),
-              cu_seqlens_q.data<int>(),
-              seq_lens.data<int>(),
-              seq_lens_encoder.data<int>(),
-              cos_emb,
-              sin_emb,
-              const_cast<DataType_*>(reinterpret_cast<const DataType_*>(cache_k_scale.get().data<T>())),
-              const_cast<DataType_*>(reinterpret_cast<const DataType_*>((cache_v_scale.get().data<T>()))),
-              nullptr,
-              nullptr,
-              max_seq_len,
-              max_blocks_per_seq,
-              num_heads,
-              block_size,
-              127.0f,
-              -127.0f,
-              kv_num_heads,
-              rope_3d,
-              rms_norm_eps);
-    } else if (cache_quant_type_str == "cache_int4_zp") {
-      append_decode_cache_int4_rope(
-          reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-          key_cache_out->data<uint8_t>(),
-          value_cache_out->data<uint8_t>(),
-          reinterpret_cast<DataType_*>(const_cast<T*>(qkv_out->data<T>())),
-          block_tables.data<int>(),
-          batch_id_per_token.data<int>(),
-          cu_seqlens_q.data<int>(),
-          seq_lens.data<int>(),
-          seq_lens_encoder.data<int>(),
-          cos_emb,
-          sin_emb,
-          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-          qkv_biases ? reinterpret_cast<DataType_*>(
-                          const_cast<T*>(qkv_biases.get().data<T>()))
-                    : nullptr,
-          cache_k_scale ? reinterpret_cast<DataType_*>(
-                              const_cast<T*>(cache_k_scale.get().data<T>()))
-                        : nullptr,
-          cache_v_scale ? reinterpret_cast<DataType_*>(
-                              const_cast<T*>(cache_v_scale.get().data<T>()))
-                        : nullptr,
-          cache_k_zp ? reinterpret_cast<DataType_*>(
-                          const_cast<T*>(cache_k_zp.get().data<T>()))
-                    : nullptr,
-          cache_v_zp ? reinterpret_cast<DataType_*>(
-                          const_cast<T*>(cache_v_zp.get().data<T>()))
-                    : nullptr,
-          max_seq_len,
-          max_blocks_per_seq,
-          num_heads,
-          kv_num_heads,
-          dim_head,
-          block_size,
-          bsz,
-          stream,
-          use_neox_rotary_style,
-          rope_3d);
-    } else {
-      PD_THROW(
-          "cache_quant_type_str should be one of [none, cache_int8, cache_fp8 "
-          "cache_int4_zp]");
-    }
+    PD_THROW(
+        "cache_quant_type_str should be one of [none, cache_int8, cache_fp8 "
+        "cache_int4_zp]");
   }
 }
 
@@ -826,8 +650,8 @@ template void DecoderWriteCacheWithRoPEKernel<paddle::bfloat16, int>(
               // kv_num_heads, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& qkv_out_scales,
@@ -843,10 +667,7 @@ template void DecoderWriteCacheWithRoPEKernel<paddle::bfloat16, int>(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps);
+    paddle::Tensor* value_cache_out);
 
 template void
 DecoderWriteCacheWithRoPEKernel<paddle::bfloat16, paddle::bfloat16>(
@@ -856,8 +677,8 @@ DecoderWriteCacheWithRoPEKernel<paddle::bfloat16, paddle::bfloat16>(
               // kv_num_heads, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& qkv_out_scales,
@@ -873,10 +694,7 @@ DecoderWriteCacheWithRoPEKernel<paddle::bfloat16, paddle::bfloat16>(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps);
+    paddle::Tensor* value_cache_out);
 
 template void DecoderWriteCacheWithRoPEKernel<paddle::float16, int>(
     const AppendAttnMetaData& meta_data,
@@ -885,8 +703,8 @@ template void DecoderWriteCacheWithRoPEKernel<paddle::float16, int>(
               // kv_num_heads, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& qkv_out_scales,
@@ -902,10 +720,7 @@ template void DecoderWriteCacheWithRoPEKernel<paddle::float16, int>(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps);
+    paddle::Tensor* value_cache_out);
 
 template void DecoderWriteCacheWithRoPEKernel<paddle::float16, paddle::float16>(
     const AppendAttnMetaData& meta_data,
@@ -914,8 +729,8 @@ template void DecoderWriteCacheWithRoPEKernel<paddle::float16, paddle::float16>(
               // kv_num_heads, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& qkv_out_scales,
@@ -931,7 +746,4 @@ template void DecoderWriteCacheWithRoPEKernel<paddle::float16, paddle::float16>(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps);
+    paddle::Tensor* value_cache_out);

custom_ops/gpu_ops/append_attn/decoder_write_cache_with_rope_kernel.h

@@ -23,8 +23,8 @@ void DecoderWriteCacheWithRoPEKernel(
               // kv_num_heads, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& qkv_out_scales,
@@ -40,6 +40,4 @@ void DecoderWriteCacheWithRoPEKernel(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,    const float rms_norm_eps);
+    paddle::Tensor* value_cache_out);

custom_ops/gpu_ops/append_attn/encoder_write_cache_with_rope_kernel.h

@@ -25,8 +25,8 @@ void EncoderWriteCacheWithRopeKernel(
     const paddle::Tensor& seq_lens_this_time,
     const paddle::Tensor& seq_lens_encoder,
     const paddle::Tensor& seq_lens_decoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids,
@@ -46,10 +46,7 @@ void EncoderWriteCacheWithRopeKernel(
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps) {
+    paddle::Tensor* value_cache_out) {
   auto token_num = meta_data.token_nums;
   auto num_heads = meta_data.q_num_heads;
   auto kv_num_heads = meta_data.kv_num_heads;
@@ -59,16 +56,33 @@ void EncoderWriteCacheWithRopeKernel(
     is_scale_channel_wise = true;
   }
 
-  if (q_norm_weight && k_norm_weight) {
-    if (num_heads != kv_num_heads && !is_scale_channel_wise && !use_neox_style) {
-      gqa_rotary_qk_norm_variable(
+  if (num_heads == kv_num_heads) {
+    rotary_qk_variable(
         qkv_out->data<T>(),
         qkv.data<QKV_TYPE>(),
         qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
         qkv_biases ? qkv_biases.get().data<T>() : nullptr,
         rotary_embs.get().data<float>(),
-        batch_id_per_token.data<int>(),
-        cu_seqlens_q.data<int>(),
+        padding_offsets.data<int>(),
+        seq_lens_encoder.data<int>(),
+        seq_lens_decoder.data<int>(),
+        token_num,
+        num_heads,
+        max_seq_len,
+        rotary_embs.get().dims()[2],
+        head_dim,
+        stream,
+        use_neox_style,
+        rope_3d);
+  } else {
+    if (!is_scale_channel_wise) {
+      gqa_rotary_qk_variable(
+        qkv_out->data<T>(),
+        qkv.data<QKV_TYPE>(),
+        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+        qkv_biases ? qkv_biases.get().data<T>() : nullptr,
+        rotary_embs.get().data<float>(),
+        padding_offsets.data<int>(),
         seq_lens_encoder.data<int>(),
         seq_lens_decoder.data<int>(),
         token_num,
@@ -79,95 +93,44 @@ void EncoderWriteCacheWithRopeKernel(
         head_dim,
         stream,
         use_neox_style,
-        rope_3d,
-        q_norm_weight ? q_norm_weight.get().data<float>() : nullptr,
-        k_norm_weight ? k_norm_weight.get().data<float>() : nullptr,
-        rms_norm_eps);
+        rope_3d);
     } else {
-      PD_THROW(
-          "gqa_rotary_qk_norm_variable only support gqa mode. channel wise scale and neox style are not supported");
+      gqa_rotary_qk_quant_variable(
+        qkv_out->data<T>(),
+        qkv.data<QKV_TYPE>(),
+        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+        qkv_biases ? qkv_biases.get().data<T>() : nullptr,
+        cache_k_scale ? cache_k_scale.get().data<T>() : nullptr,
+        cache_v_scale ? cache_v_scale.get().data<T>() : nullptr,
+        rotary_embs.get().data<float>(),
+        padding_offsets.data<int>(),
+        seq_lens_encoder.data<int>(),
+        seq_lens_decoder.data<int>(),
+        token_num,
+        num_heads,
+        kv_num_heads,
+        max_seq_len,
+        rotary_embs.get().dims()[2],
+        head_dim,
+        stream,
+        use_neox_style,
+        rope_3d);
     }
-  } else {
-    if (num_heads == kv_num_heads) {
-      rotary_qk_variable(
-          qkv_out->data<T>(),
-          qkv.data<QKV_TYPE>(),
-          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-          qkv_biases ? qkv_biases.get().data<T>() : nullptr,
-          rotary_embs.get().data<float>(),
-          batch_id_per_token.data<int>(),
-          cu_seqlens_q.data<int>(),
-          seq_lens_encoder.data<int>(),
-          seq_lens_decoder.data<int>(),
-          token_num,
-          num_heads,
-          max_seq_len,
-          rotary_embs.get().dims()[2],
-          head_dim,
-          stream,
-          use_neox_style,
-          rope_3d);
-    } else {
-      if (!is_scale_channel_wise) {
-        gqa_rotary_qk_variable(
-          qkv_out->data<T>(),
-          qkv.data<QKV_TYPE>(),
-          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-          qkv_biases ? qkv_biases.get().data<T>() : nullptr,
-          rotary_embs.get().data<float>(),
-          batch_id_per_token.data<int>(),
-          cu_seqlens_q.data<int>(),
-          seq_lens_encoder.data<int>(),
-          seq_lens_decoder.data<int>(),
-          token_num,
-          num_heads,
-          kv_num_heads,
-          max_seq_len,
-          rope_3d ? rotary_embs.get().dims()[3] : rotary_embs.get().dims()[2],
-          head_dim,
-          stream,
-          use_neox_style,
-          rope_3d);
-      } else {
-        gqa_rotary_qk_quant_variable(
-          qkv_out->data<T>(),
-          qkv.data<QKV_TYPE>(),
-          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-          qkv_biases ? qkv_biases.get().data<T>() : nullptr,
-          cache_k_scale ? cache_k_scale.get().data<T>() : nullptr,
-          cache_v_scale ? cache_v_scale.get().data<T>() : nullptr,
-          rotary_embs.get().data<float>(),
-          batch_id_per_token.data<int>(),
-          cu_seqlens_q.data<int>(),
-          seq_lens_encoder.data<int>(),
-          seq_lens_decoder.data<int>(),
-          token_num,
-          num_heads,
-          kv_num_heads,
-          max_seq_len,
-          rotary_embs.get().dims()[2],
-          head_dim,
-          stream,
-          use_neox_style,
-          rope_3d);
-      }
 
-    }
   }
   const uint32_t block_size = meta_data.block_size;
   if (cache_quant_type_str == "none") {
     CascadeAppendWriteCacheKVQKV<T>(meta_data,
                                     *qkv_out,
                                     block_tables,
-                                    batch_id_per_token,
-                                    cu_seqlens_q,
+                                    padding_offsets,
                                     seq_lens_encoder,
                                     seq_lens_decoder,
                                     max_seq_len,
                                     stream,
                                     key_cache_out,
                                     value_cache_out);
-  } else if (cache_quant_type_str == "cache_int8" or cache_quant_type_str == "cache_fp8" or cache_quant_type_str == "block_wise_fp8") {
+  } else if (cache_quant_type_str == "cache_int8" or cache_quant_type_str == "cache_fp8") {
     DISPATCH_HEAD_DIM(
         head_dim, HEAD_DIM, {DISPATCH_BLOCK_SIZE(block_size, BLOCK_SIZE, {
           CascadeAppendWriteCacheKVC8QKV<T, HEAD_DIM, BLOCK_SIZE>(
@@ -179,15 +142,15 @@ void EncoderWriteCacheWithRopeKernel(
               cache_v_scale.get(),
               seq_lens_this_time,
               seq_lens_decoder,
-              batch_id_per_token,
-              cu_seqlens_q,
+              padding_offsets,
+              cum_offsets,
               block_tables,
               batch_ids,
               tile_ids,
               num_blocks,
               max_seq_len,
               is_scale_channel_wise,
-              cache_quant_type_str,
+              cache_quant_type_str == "cache_fp8",
               stream,
               key_cache_out,
               value_cache_out);
@@ -206,8 +169,8 @@ void EncoderWriteCacheWithRopeKernel(
               cache_v_zp.get(),
               seq_lens_this_time,
               seq_lens_decoder,
-              batch_id_per_token,
-              cu_seqlens_q,
+              padding_offsets,
+              cum_offsets,
               block_tables,
               batch_ids,
               tile_ids,

custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu

@@ -194,26 +194,23 @@ get_max_len_kv_ernel(int *max_seq_lens_out, const int *seq_lens_this_time,
 std::vector<paddle::Tensor> GetBlockShapeAndSplitKVBlock(
     const paddle::Tensor &seq_lens_encoder,
     const paddle::Tensor &seq_lens_decoder,
-    const paddle::Tensor &seq_lens_this_time,
-    paddle::Tensor &decoder_batch_ids,          // Inplace
-    paddle::Tensor &decoder_tile_ids_per_batch, // Inplace
-    paddle::Tensor &decoder_num_blocks_x_cpu,   // Inplace, Pinned Memory
-    paddle::Tensor &max_len_tensor_cpu,         // Inplace, Pinned Memory
-    const int encoder_block_shape_q,
-    const int decoder_block_shape_q,
-    const int group_size,
-    const int block_size,
-    const int decoder_step_token_num)
-{
+    const paddle::Tensor &seq_lens_this_time, const paddle::Tensor &cum_offsets,
+    const int encoder_block_shape_q, const int decoder_block_shape_q,
+    const int group_size, const int block_size,
+    const int decoder_step_token_num) {
   auto stream = seq_lens_encoder.stream();
-  int bsz = seq_lens_this_time.shape()[0];
-
-  paddle::Tensor max_len_tensor_gpu = GetEmptyTensor({max_len_tensor_cpu.shape()[0]}, paddle::DataType::INT32, seq_lens_this_time.place());
+  int bsz = cum_offsets.shape()[0];
+  auto max_len_tensor =
+      GetEmptyTensor({8}, paddle::DataType::INT32, seq_lens_encoder.place());
   GetMaxLen(seq_lens_decoder, seq_lens_this_time, seq_lens_encoder,
-            max_len_tensor_gpu, bsz);
-  max_len_tensor_cpu.copy_(max_len_tensor_gpu, max_len_tensor_cpu.place(), false);
+            max_len_tensor, bsz);
 
-  auto max_len_cpu_ptr = max_len_tensor_cpu.data<int>();
+  // max_len_this_time, max_enc_len_this_time, max_dec_len_this_time,
+  // max_enc_dec_len_this_time, max_just_dec_len_this_time,
+  // max_just_dec_merged_len_this_time, max_system_len,
+  // max_just_dec_len_without_system
+  auto max_len_cpu = max_len_tensor.copy_to(paddle::CPUPlace(), false);
+  auto max_len_cpu_ptr = max_len_cpu.data<int>();
   int max_len_this_time = max_len_cpu_ptr[0];
   int max_enc_len_this_time = max_len_cpu_ptr[1];
   int max_dec_len_this_time = max_len_cpu_ptr[2];
@@ -225,11 +222,14 @@ std::vector<paddle::Tensor> GetBlockShapeAndSplitKVBlock(
 
   paddle::Tensor encoder_batch_ids;
   paddle::Tensor encoder_tile_ids_per_batch;
-  paddle::Tensor encoder_num_blocks_x_cpu;  /*cpu*/
+  paddle::Tensor encoder_num_blocks_x_cpu; /*cpu*/
   paddle::Tensor kv_batch_ids;
   paddle::Tensor kv_tile_ids_per_batch;
-  paddle::Tensor kv_num_blocks_x_cpu;       /*cpu*/
-  paddle::Tensor max_len_kv_cpu;            /*cpu*/
+  paddle::Tensor kv_num_blocks_x_cpu; /*cpu*/
+  paddle::Tensor decoder_batch_ids;
+  paddle::Tensor decoder_tile_ids_per_batch;
+  paddle::Tensor decoder_num_blocks_x_cpu; /*cpu*/
+  paddle::Tensor max_len_kv_cpu;           /*cpu*/
 
   auto max_len_kv =
       GetEmptyTensor({1}, paddle::DataType::INT32, seq_lens_decoder.place());
@@ -289,66 +289,97 @@ std::vector<paddle::Tensor> GetBlockShapeAndSplitKVBlock(
     kv_tile_ids_per_batch =
         GetEmptyTensor({0}, paddle::DataType::INT32, seq_lens_encoder.place());
     kv_num_blocks_x_cpu =
-        GetEmptyTensor({0}, paddle::DataType::INT32, paddle::CPUPlace());
+        GetEmptyTensor({0}, paddle::DataType::INT32, seq_lens_encoder.place());
   }
-
   if (max_just_dec_len_this_time > 0) {
-    // Clear buffer
-    const uint32_t decoder_max_tile_size_per_bs_q = div_up((decoder_step_token_num * group_size), decoder_block_shape_q);
-    const uint32_t decoder_batch_shape = bsz * decoder_max_tile_size_per_bs_q;
-    PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(decoder_batch_ids.data<int>(), 0, decoder_batch_shape * sizeof(int32_t), stream));
-    PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(decoder_tile_ids_per_batch.data<int>(), 0, decoder_batch_shape * sizeof(int32_t), stream));
-    PADDLE_ENFORCE_GPU_SUCCESS(cudaMemsetAsync(decoder_num_blocks_x_cpu.data<int>(), 0, sizeof(int32_t), stream));
+    const uint32_t decoder_max_tile_size_per_bs_q =
+        div_up((decoder_step_token_num * group_size), decoder_block_shape_q);
 
+    decoder_batch_ids =
+        GetEmptyTensor({bsz * decoder_max_tile_size_per_bs_q},
+                       paddle::DataType::INT32, seq_lens_encoder.place());
+    decoder_tile_ids_per_batch =
+        GetEmptyTensor({bsz * decoder_max_tile_size_per_bs_q},
+                       paddle::DataType::INT32, seq_lens_encoder.place());
     auto decoder_num_blocks_x =
         GetEmptyTensor({1}, paddle::DataType::INT32, seq_lens_encoder.place());
     split_q_block<<<1, 32, 0, stream>>>(
-        seq_lens_this_time.data<int>(),
-        seq_lens_encoder.data<int>(),
-        decoder_batch_ids.data<int>(),
-        decoder_tile_ids_per_batch.data<int>(),
-        decoder_num_blocks_x.data<int>(),
-        bsz,
-        decoder_block_shape_q,
+        seq_lens_this_time.data<int>(), seq_lens_encoder.data<int>(),
+        decoder_batch_ids.data<int>(), decoder_tile_ids_per_batch.data<int>(),
+        decoder_num_blocks_x.data<int>(), bsz, decoder_block_shape_q,
         group_size);
-    decoder_num_blocks_x_cpu.copy_(decoder_num_blocks_x, decoder_num_blocks_x_cpu.place(), false);
+    decoder_num_blocks_x_cpu =
+        decoder_num_blocks_x.copy_to(paddle::CPUPlace(), false);
+  } else {
+    decoder_batch_ids =
+        GetEmptyTensor({0}, paddle::DataType::INT32, seq_lens_encoder.place());
+    decoder_tile_ids_per_batch =
+        GetEmptyTensor({0}, paddle::DataType::INT32, seq_lens_encoder.place());
+    decoder_num_blocks_x_cpu =
+        GetEmptyTensor({0}, paddle::DataType::INT32, paddle::CPUPlace());
   }
 
+  return {encoder_batch_ids,
+          encoder_tile_ids_per_batch,
+          encoder_num_blocks_x_cpu, /*cpu*/
+          kv_batch_ids,
+          kv_tile_ids_per_batch,
+          kv_num_blocks_x_cpu, /*cpu*/
+          decoder_batch_ids,
+          decoder_tile_ids_per_batch,
+          decoder_num_blocks_x_cpu, /*cpu*/
+          max_len_kv_cpu /*cpu*/,
+          max_len_cpu};
+}
+
+std::vector<paddle::DataType> GetBlockShapeAndSplitKVBlockInferDtype(
+    const paddle::DataType &seq_lens_encoder_dtype,
+    const paddle::DataType &seq_lens_decoder_dtype,
+    const paddle::DataType &seq_lens_this_time_dtype,
+    const paddle::DataType &cum_offsets_dtype) {
   return {
-    encoder_batch_ids,
-    encoder_tile_ids_per_batch,
-    encoder_num_blocks_x_cpu, /*cpu*/
-    kv_batch_ids,
-    kv_tile_ids_per_batch,
-    kv_num_blocks_x_cpu,      /*cpu*/
-    max_len_kv_cpu,           /*cpu*/
-  };
+      paddle::DataType::INT32, paddle::DataType::INT32, paddle::DataType::INT32,
+      paddle::DataType::INT32, paddle::DataType::INT32, paddle::DataType::INT32,
+      paddle::DataType::INT32, paddle::DataType::INT32, paddle::DataType::INT32,
+      paddle::DataType::INT32, paddle::DataType::INT32};
+}
+
+std::vector<std::vector<int64_t>> GetBlockShapeAndSplitKVBlockInferShape(
+    const std::vector<int64_t> &seq_lens_encoder_shape,
+    const std::vector<int64_t> &seq_lens_decoder_shape,
+    const std::vector<int64_t> &seq_lens_this_time_shape,
+    const std::vector<int64_t> &cum_offsets_shape) {
+  std::vector<int64_t> dynamic_shape = {-1};
+
+  return {dynamic_shape,
+          dynamic_shape,
+          {1},
+          dynamic_shape,
+          dynamic_shape,
+          {1},
+          dynamic_shape,
+          dynamic_shape,
+          {1},
+          {1},
+          {8}};
 }
 
 PD_BUILD_STATIC_OP(get_block_shape_and_split_kv_block)
-    .Inputs({
-      "seq_lens_encoder",
-      "seq_lens_decoder",
-      "seq_lens_this_time",
-      "decoder_batch_ids",
-      "decoder_tile_ids_per_batch",
-      "decoder_num_blocks_x_cpu",
-      "max_len_tensor_cpu"
-    })
-    .Outputs({
-      paddle::Optional("encoder_batch_ids"),
-      paddle::Optional("encoder_tile_ids_per_batch"),
-      paddle::Optional("encoder_num_blocks_x_cpu"),
-      paddle::Optional("kv_batch_ids"),
-      paddle::Optional("kv_tile_ids_per_batch"),
-      paddle::Optional("kv_num_blocks_x_cpu"),
-      "max_len_kv_cpu"
-    })
-    .Attrs({
-      "encoder_block_shape_q: int",
-      "decoder_block_shape_q: int",
-      "group_size: int",
-      "block_size: int",
-      "decoder_step_token_num: int"
-    })
-    .SetKernelFn(PD_KERNEL(GetBlockShapeAndSplitKVBlock));
+    .Inputs({"seq_lens_encoder", "seq_lens_decoder", "seq_lens_this_time",
+             "cum_offsets"})
+    .Outputs({paddle::Optional("encoder_batch_ids"),
+              paddle::Optional("encoder_tile_ids_per_batch"),
+              paddle::Optional("encoder_num_blocks"),
+              paddle::Optional("kv_batch_ids"),
+              paddle::Optional("kv_tile_ids_per_batch"),
+              paddle::Optional("kv_num_blocks"),
+              paddle::Optional("decoder_batch_ids"),
+              paddle::Optional("decoder_tile_ids_per_batch"),
+              paddle::Optional("decoder_num_blocks"),
+              paddle::Optional("max_len_kv"), "set_max_lengths"})
+    .Attrs({"encoder_block_shape_q: int", "decoder_block_shape_q: int",
+            "group_size: int", "block_size: int",
+            "decoder_step_token_num: int"})
+    .SetKernelFn(PD_KERNEL(GetBlockShapeAndSplitKVBlock))
+    .SetInferShapeFn(PD_INFER_SHAPE(GetBlockShapeAndSplitKVBlockInferShape))
+    .SetInferDtypeFn(PD_INFER_DTYPE(GetBlockShapeAndSplitKVBlockInferDtype));

custom_ops/gpu_ops/append_attn/gqa_rope_write_cache.cu

@@ -16,6 +16,7 @@
 #include "paddle/extension.h"
 #include "paddle/phi/core/memory/memcpy.h"
 #include "encoder_write_cache_with_rope_impl.cuh"
+#include "paddle/phi/kernels/gpu/flash_attn_v3_kernel.h"
 #include "paddle/phi/backends/context_pool.h"
 #include "remote_cache_kv_ipc.h"
 
@@ -24,8 +25,7 @@ __global__ void GQAVariableLengthRotarySplitKernel(
     const T *qkv,
     const float *cos_emb,
     const float *sin_emb,
-    const int *batch_id_per_token,
-    const int *cu_seqlens_q,
+    const int *padding_offsets,
     const int *seq_lens,
     const int *seq_lens_decoder,
     const int *cu_seqlens_k,
@@ -37,8 +37,7 @@ __global__ void GQAVariableLengthRotarySplitKernel(
     const int q_num_head,
     const int kv_num_head,
     const int seq_len,
-    const int last_dim,
-    const bool rope_3d) {
+    const int last_dim) {
   using LoadT = AlignedVector<T, VecSize>;
   constexpr int HalfVecSize = VecSize / 2;
   using LoadEmbT = AlignedVector<float, HalfVecSize>;
@@ -53,17 +52,17 @@ __global__ void GQAVariableLengthRotarySplitKernel(
        linear_index < elem_cnt;
        linear_index += step) {
     const int token_idx = linear_index / offset;
-    const int ori_bi = batch_id_per_token[token_idx];
+    const int ori_token_idx = token_idx + padding_offsets[token_idx];
+    const int ori_bi = ori_token_idx / seq_len;
     if (seq_lens[ori_bi] == 0) continue;
     const int bias = linear_index % offset;
     const int hi = bias / last_dim;
     const int h_bias = bias % last_dim;
 
-    const int ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
+    const int ori_seq_id = ori_token_idx % seq_len + seq_lens_decoder[ori_bi];
     const int kv_write_idx = cu_seqlens_k[ori_bi] + ori_seq_id;
 
     const int64_t emb_idx = ori_seq_id * half_lastdim + h_bias / 2;
-    int64_t new_emb_idx = rope_3d ? emb_idx + ori_bi * last_dim * seq_len : emb_idx;
     const int64_t base_idx =
         token_idx * (q_num_head + 2 * kv_num_head) * last_dim + hi * last_dim +
         h_bias;
@@ -82,8 +81,8 @@ __global__ void GQAVariableLengthRotarySplitKernel(
     Load<T, VecSize>(&qkv[base_idx], &src_vec);
     // do rope
     if (hi < q_num_head + kv_num_head) {
-      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < HalfVecSize; i++) {
         const float input_left = static_cast<float>(src_vec[2 * i]);
@@ -109,10 +108,9 @@ void gqa_rotary_qk_split_variable(
     T *v,
     const T *qkv_input,
     const float *rotary_emb,  // [2, 1, 1, seq_len, dim_head / 2]
-    const int *batch_id_per_token,
+    const int *padding_offsets,
     const int *seq_lens_encoder,
     const int *seq_lens_decoder,
-    const int *cu_seqlens_q,
     const int *cu_seqlens_k,
     const int token_num,
     const int num_heads,
@@ -120,7 +118,6 @@ void gqa_rotary_qk_split_variable(
     const int seq_len,
     const int input_output_len,
     const int dim_head,
-    const bool rope_3d,
     const cudaStream_t &stream) {
   int64_t elem_nums = token_num * (num_heads + 2 * kv_num_heads) * dim_head;
   constexpr int PackSize = 16 / sizeof(T);
@@ -136,8 +133,7 @@ void gqa_rotary_qk_split_variable(
             qkv_input,
             cos_emb,
             sin_emb,
-            batch_id_per_token,
-            cu_seqlens_q,
+            padding_offsets,
             seq_lens_encoder,
             seq_lens_decoder,
             cu_seqlens_k,
@@ -149,183 +145,7 @@ void gqa_rotary_qk_split_variable(
             num_heads,
             kv_num_heads,
             seq_len,
-            dim_head,
-            rope_3d);
-}
-
-template <typename T,
-          typename CacheT,
-          uint32_t HEAD_DIM,
-          uint32_t BLOCK_SIZE,
-          uint32_t NUM_WARPS=4>
-__global__ void append_cache_kv_c16(
-    const T *__restrict__ cache_k,
-    const T *__restrict__ cache_v,
-    T *__restrict__ k_out,
-    T *__restrict__ v_out,
-    const int *__restrict__ seq_lens_this_time,
-    const int *__restrict__ seq_lens_decoder,
-    const int *__restrict__ cu_seqlens_k,
-    const int *__restrict__ block_tables,
-    const int *batch_ids,
-    const int *tile_ids_per_batch,
-    const int max_blocks_per_seq,
-    const int kv_num_heads) {
-  // start_kv_idx: start kv_idx current block
-  // batch_id：block's batch_id
-  // TODO: 1.scale preload 2.frag_dq_T reuse 3.pipeline 4.store aligned 5.cacheT with template（int8/fp8)
-  const uint32_t tile_idx = blockIdx.x, kv_head_idx = blockIdx.z;
-  const uint32_t tid = threadIdx.x, wid = threadIdx.y;
-
-  const uint32_t batch_id = batch_ids[tile_idx];
-  const uint32_t start_kv_idx = tile_ids_per_batch[tile_idx] * BLOCK_SIZE;
-  const uint32_t end_idx = seq_lens_decoder[batch_id] - start_kv_idx;
-  if (seq_lens_this_time[batch_id] <= 0) {
-    return;
-  }
-
-  const int *cur_block_table = block_tables + batch_id * max_blocks_per_seq;
-  uint32_t block_id = cur_block_table[start_kv_idx / BLOCK_SIZE];
-  // cache_kv idx
-  uint32_t kv_h_stride = BLOCK_SIZE * HEAD_DIM;
-  uint32_t block_stride = kv_num_heads * kv_h_stride;
-  const CacheT *cur_cache_k = cache_k + block_id * block_stride + kv_head_idx * kv_h_stride;
-  const CacheT *cur_cache_v = cache_v + block_id * block_stride + kv_head_idx * kv_h_stride;
-
-  // k_out v_out idx
-  uint32_t kv_t_stride = kv_num_heads * HEAD_DIM;
-  T *k_write_ptr = k_out + (cu_seqlens_k[batch_id] + start_kv_idx) * kv_t_stride;
-  T *v_write_ptr = v_out + (cu_seqlens_k[batch_id] + start_kv_idx) * kv_t_stride;
-
-  uint32_t kv_frag[4];
-  T *frag_dq_T = reinterpret_cast<T *>(kv_frag);
-
-  constexpr uint32_t num_vecs_per_head =
-      HEAD_DIM / num_elems_per_128b<CacheT>();
-  constexpr uint32_t inv_kv_stride = 8 / num_vecs_per_head;
-
-  extern __shared__ uint8_t smem[];
-  smem_t k_smem(smem);
-  uint32_t k_smem_offset_w = smem_t::get_permuted_offset<num_vecs_per_head, inv_kv_stride>(
-      wid * 4 + tid / 8, tid % 8);  // 4 * 4 per warp
-
-  uint32_t k_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head, inv_kv_stride>(
-      wid * 16 + 8 * (tid / 16) + tid % 8, (tid % 16) / 8);
-
-  uint32_t k_read_idx = (wid * 4 + tid / 8) * HEAD_DIM +
-                          tid % 8 * num_elems_per_128b<CacheT>();
-
-  // load k_smem 64 rows 128 cols
-  for (int fz = 0; fz < 4; fz++) { // 4 rows pre warp once, 16 rows all 4 warps once, need 4 iter
-    for (int fy = 0; fy < 2; fy++) { // 8 * 128b = 64 * bf16 noce, need 2 iter
-      k_smem.load_128b_async<SharedMemFillMode::kNoFill>(
-            k_smem_offset_w, cur_cache_k + k_read_idx, end_idx > 0);
-      k_smem_offset_w =
-            k_smem.advance_offset_by_column<8, num_vecs_per_head>(k_smem_offset_w, fy);
-      k_read_idx += 8 * num_elems_per_128b<CacheT>();
-    }
-    k_smem_offset_w =
-          k_smem.advance_offset_by_row<4 * NUM_WARPS, num_vecs_per_head>(k_smem_offset_w) - 16;
-    k_read_idx += 4 * NUM_WARPS * HEAD_DIM - 16 * num_elems_per_128b<CacheT>();
-  }
-  commit_group();
-  wait_group<0>();
-  __syncthreads();
-
-  // deal k_smem 64 rows 128 cols
-  for (int fz = 0; fz < 1; fz++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 1 iter
-    uint32_t row_idx = wid * 16 + tid / 4;
-    for (int fy = 0; fy < 8; fy++) { // 2 * 128b = 16 * bf16 noce, need 8 iter
-      uint32_t col_idx = fy * 16 + tid % 4 * 2;
-      k_smem.ldmatrix_m8n8x4(k_smem_offset_r, kv_frag);
-      // layout
-      /***
-        r0c0,r0c1, r0c8,r0c9
-        r8c0,r8c1, r8c8,r8c9
-      ***/
-      T *k_tile_ptr0 = k_write_ptr + row_idx * kv_t_stride + kv_head_idx * HEAD_DIM + col_idx;
-      T *k_tile_ptr1 = k_tile_ptr0 + 8 * kv_t_stride;
-
-      if (row_idx < end_idx) {
-        k_tile_ptr0[0] = frag_dq_T[0];
-        k_tile_ptr0[1] = frag_dq_T[1];
-        k_tile_ptr0[8] = frag_dq_T[2];
-        k_tile_ptr0[9] = frag_dq_T[3];
-      }
-
-      if (row_idx + 8 < end_idx) {
-        k_tile_ptr1[0] = frag_dq_T[4];
-        k_tile_ptr1[1] = frag_dq_T[5];
-        k_tile_ptr1[8] = frag_dq_T[6];
-        k_tile_ptr1[9] = frag_dq_T[7];
-      }
-      k_smem_offset_r = k_smem.advance_offset_by_column<2, num_vecs_per_head>(
-        k_smem_offset_r, fy);
-    }
-    k_smem_offset_r =
-      k_smem.advance_offset_by_row<16 * NUM_WARPS, num_vecs_per_head>(k_smem_offset_r) - 16;
-  }
-
-  // ================v================
-  smem_t v_smem(smem + BLOCK_SIZE * HEAD_DIM * sizeof(CacheT));
-  uint32_t v_smem_offset_w = smem_t::get_permuted_offset<num_vecs_per_head, inv_kv_stride>(
-      wid * 4 + tid / 8, tid % 8);  // 4 * 4 per warp
-  uint32_t v_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head, inv_kv_stride>(
-      wid * 16 + 8 * (tid / 16) + tid % 8, (tid % 16) / 8);
-
-  uint32_t v_read_idx = (wid * 4 + tid / 8) * HEAD_DIM +
-                          tid % 8 * num_elems_per_128b<CacheT>();
-
-  // load v_smem 64 rows 128 cols
-  for (int fz = 0; fz < 4; fz++) { // // 4 rows pre warp once, 16 rows all 4 warps once, need 4 iter
-    for (int fy = 0; fy < 2; fy++) { // 8 * 128b = 64 * bf16 noce, need 2 iter
-      v_smem.load_128b_async<SharedMemFillMode::kNoFill>(
-            v_smem_offset_w, cur_cache_v + v_read_idx, end_idx > 0);
-      v_smem_offset_w =
-            v_smem.advance_offset_by_column<8, num_vecs_per_head>(v_smem_offset_w, fy);
-      v_read_idx += 8 * num_elems_per_128b<CacheT>();
-    }
-    v_smem_offset_w =
-          v_smem.advance_offset_by_row<4 * NUM_WARPS, num_vecs_per_head>(v_smem_offset_w) - 16;
-    v_read_idx += 4 * NUM_WARPS * HEAD_DIM - 16 * num_elems_per_128b<CacheT>();
-  }
-  commit_group();
-  wait_group<0>();
-  __syncthreads();
-
-  // deal v_smem 64 rows 128 cols
-  for (int fz = 0; fz < 1; fz++) { //  16 rows pre warp once, 64 rows all 4 warps once, need 1 iter
-    uint32_t row_idx = wid * 16 + tid / 4;
-    for (int fy = 0; fy < 8; fy++) { // 2 * 128b = 16 * bf16 noce, need 8 iter
-      uint32_t col_idx = fy * 16 + tid % 4 * 2;
-      v_smem.ldmatrix_m8n8x4(v_smem_offset_r, kv_frag);
-      // layout
-      /***
-        r0c0,r0c1, r0c8,r0c9
-        r8c0,r8c1, r8c8,r8c9
-      ***/
-      T *v_tile_ptr0 = v_write_ptr + row_idx * kv_t_stride + kv_head_idx * HEAD_DIM + col_idx;
-      T *v_tile_ptr1 = v_tile_ptr0 + 8 * kv_t_stride;
-
-      if (row_idx < end_idx) {
-        v_tile_ptr0[0] = frag_dq_T[0];
-        v_tile_ptr0[1] = frag_dq_T[1];
-        v_tile_ptr0[8] = frag_dq_T[2];
-        v_tile_ptr0[9] = frag_dq_T[3];
-      }
-
-      if (row_idx + 8 < end_idx) {
-        v_tile_ptr1[0] = frag_dq_T[4];
-        v_tile_ptr1[1] = frag_dq_T[5];
-        v_tile_ptr1[8] = frag_dq_T[6];
-        v_tile_ptr1[9] = frag_dq_T[7];
-      }
-      v_smem_offset_r = v_smem.advance_offset_by_column<2, num_vecs_per_head>(
-        v_smem_offset_r, fy);
-    }
-    v_smem_offset_r =
-      v_smem.advance_offset_by_row<16 * NUM_WARPS, num_vecs_per_head>(v_smem_offset_r) - 16;
-  }
+            dim_head);
 }
 
 template <typename T,
@@ -334,7 +154,7 @@ template <typename T,
           uint32_t BLOCK_SIZE,
           uint32_t NUM_WARPS=4,
           bool IS_FP8=false>
-__global__ void append_cache_kv_c8(
+__global__ void append_dequant_cache_kv_c8(
     const CacheT *__restrict__ cache_k,
     const CacheT *__restrict__ cache_v,
     T *__restrict__ k_out,
@@ -349,16 +169,16 @@ __global__ void append_cache_kv_c8(
     const int *tile_ids_per_batch,
     const int max_blocks_per_seq,
     const int kv_num_heads) {
-  // start_kv_idx: start kv_idx current block
-  // batch_id：block's batch_id
-  // TODO: 1.scale preload 2.frag_dq_T reuse 3.pipeline 4.store aligned 5.cacheT with template（int8/fp8)
+  // start_kv_idx: 每个block的起始kv_idx
+  // batch_id：每个block属于的batch
+  // TODO: 1.scale预取 2.frag_dq_T复用 3.流水线编排 4.store访存合并 5.cacheT支持（int8/fp8)
   const uint32_t tile_idx = blockIdx.x, kv_head_idx = blockIdx.z;
   const uint32_t tid = threadIdx.x, wid = threadIdx.y;
 
   const uint32_t batch_id = batch_ids[tile_idx];
   const uint32_t start_kv_idx = tile_ids_per_batch[tile_idx] * BLOCK_SIZE;
   const uint32_t end_idx = seq_lens_decoder[batch_id] - start_kv_idx;
-  if (seq_lens_this_time[batch_id] <= 0) {
+  if (seq_lens_this_time <= 0) {
     return;
   }
 
@@ -372,8 +192,8 @@ __global__ void append_cache_kv_c8(
 
   // k_out v_out idx
   uint32_t kv_t_stride = kv_num_heads * HEAD_DIM;
-  T *k_write_ptr = k_out + (cu_seqlens_k[batch_id] + start_kv_idx) * kv_t_stride;
-  T *v_write_ptr = v_out + (cu_seqlens_k[batch_id] + start_kv_idx) * kv_t_stride;
+  T *k_write_ptr = k_out + (cu_seqlens_k[batch_id] + start_kv_idx) * kv_t_stride; // 当前k block起始指针
+  T *v_write_ptr = v_out + (cu_seqlens_k[batch_id] + start_kv_idx) * kv_t_stride; // 当前v block起始指针
 
   uint32_t k_frag[4], v_frag[4], frag_dq[4];
   T *frag_dq_T = reinterpret_cast<T *>(frag_dq);
@@ -394,13 +214,13 @@ __global__ void append_cache_kv_c8(
 
   uint32_t k_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head_k, inv_k_stride>(
       wid * 16 + 8 * (tid / 16) + tid % 8, (tid % 16) / 8);
-
+  
   uint32_t k_read_idx = (wid * 4 + tid / 8) * HEAD_DIM +
                           tid % 8 * num_elems_per_128b<CacheT>();
 
-  // load v_smem 64 rows, 128 cols
-  for (int fz = 0; fz < 4; fz++) { // 4 rows pre warp once, 16 rows all 4 warps once, need 4 iter
-    for (int fy = 0; fy < 1; fy++) { // 8 * 128b = 128 * uint8 noce, need 1 iter
+  // load k_smem 行是64 列是128
+  for (int fz = 0; fz < 4; fz++) { // 每个warp1次4行,循环4次16行,4个warp64行
+    for (int fy = 0; fy < 1; fy++) { // 一次8个128b = 128个uint8
       k_smem.load_128b_async<SharedMemFillMode::kNoFill>(
             k_smem_offset_w, cur_cache_k + k_read_idx, end_idx > 0);
       k_smem_offset_w =
@@ -415,13 +235,13 @@ __global__ void append_cache_kv_c8(
   wait_group<0>();
   __syncthreads();
 
-  // deal k_smem 64 rows, 128 cols
-  for (int fz = 0; fz < 1; fz++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 1 iter
+  // deal k_smem 行是64 列是128
+  for (int fz = 0; fz < 1; fz++) { // 每个warp1次16行,4个warp64行
     uint32_t row_idx = wid * 16 + tid / 4;
-    for (int fy = 0; fy < 4; fy++) { // 2 * 128b = 32 * uint8 noce, need 4 iter
+    for (int fy = 0; fy < 4; fy++) { // 1次2个128b(32个uint8),4次循环8个128b（128个uint8）
       uint32_t col_idx = fy * 32 + tid % 4 * 2;
       k_smem.ldmatrix_m8n8x4(k_smem_offset_r, k_frag);
-      // layout
+      // 反量化 存储
       /***
       r0c0,r0c1,r0c8,r0c9, r8c0,r8c1,r8c8,r8c9
       r0c16,r0c17,r0c24,r0c25, r8c16,r8c17,r8c24,r8c25
@@ -431,7 +251,8 @@ __global__ void append_cache_kv_c8(
         T *k_tile_ptr1 = k_tile_ptr0 + 8 * kv_t_stride;
 
         if (row_idx < end_idx) {
-          convert_c8<T,IS_FP8>(frag_dq_T,k_frag[2 * i]); // 4 * uint8/fp8 -> 4 * T
+          convert_c8<T,IS_FP8>(frag_dq_T,k_frag[2 * i]); // 4个uint8/fp8 -> 4个T
+
           k_tile_ptr0[0] = frag_dq_T[0] * cache_k_scale;
           k_tile_ptr0[1] = frag_dq_T[1] * cache_k_scale;
           k_tile_ptr0[8] = frag_dq_T[2] * cache_k_scale;
@@ -439,7 +260,8 @@ __global__ void append_cache_kv_c8(
         }
 
         if (row_idx + 8 < end_idx) {
-          convert_c8<T,IS_FP8>(frag_dq_T + 4,k_frag[2 * i + 1]); // 4 * uint8/fp8 -> 4 * T
+          convert_c8<T,IS_FP8>(frag_dq_T + 4,k_frag[2 * i + 1]); // 4个uint8/fp8 -> 4个T
+
           k_tile_ptr1[0] = frag_dq_T[4] * cache_k_scale;
           k_tile_ptr1[1] = frag_dq_T[5] * cache_k_scale;
           k_tile_ptr1[8] = frag_dq_T[6] * cache_k_scale;
@@ -453,8 +275,8 @@ __global__ void append_cache_kv_c8(
     k_smem_offset_r =
       k_smem.advance_offset_by_row<16 * NUM_WARPS, num_vecs_per_head_k>(k_smem_offset_r) - 8;
   }
-
   // ================v================
+
   smem_t v_smem(smem + BLOCK_SIZE * HEAD_DIM * sizeof(CacheT));
   uint32_t v_smem_offset_w = smem_t::get_permuted_offset<num_vecs_per_blocksize, inv_v_stride>(
       wid * 8 + tid / 4, tid % 4);  // 4 * 8 per warp
@@ -464,9 +286,9 @@ __global__ void append_cache_kv_c8(
 
   uint32_t v_read_idx = (wid * 8 + tid / 4) * BLOCK_SIZE +
                           tid % 4 * num_elems_per_128b<CacheT>();
-  // load v_smem 128 rows 64 cols
-  for (int fy = 0; fy < 4; fy++) { // 8 rows pre warp once, 32 rows all 4 warps once, need 4 iter
-    for (int fz = 0; fz < 1; fz++) { // 4 * 128b = 64 * uint8 noce, need 1 iter
+  // load v_smem 行是128 列是64
+  for (int fy = 0; fy < 4; fy++) { // 每个warp1次8行,循环4次32行,4个warp128行
+    for (int fz = 0; fz < 1; fz++) { // 一次4个128b = 64个uint8
       v_smem.load_128b_async<SharedMemFillMode::kNoFill>(
               v_smem_offset_w, cur_cache_v + v_read_idx, end_idx > 0);
       v_smem_offset_w =
@@ -482,32 +304,42 @@ __global__ void append_cache_kv_c8(
   wait_group<0>();
   __syncthreads();
 
-  // deal v_smem 128 rows 64 cols
-  for (int fy = 0; fy < 2; fy++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 2 iter
+  // deal v_smem 行是128 列是64 row_idx是head_dim, col_idx是block_size
+  for (int fy = 0; fy < 2; fy++) { // 每个warp1次16行,循环2次32行，4个warp128行
     uint32_t dim_idx = fy * NUM_WARPS * 16 + wid * 16 + tid / 4;
-    for (int fz = 0; fz < 2; fz++) { // 2 * 128b = 32 * uint8 noce, need 2 iter
+    for (int fz = 0; fz < 2; fz++) { // 1次2个128b(32个uint8),2次循环4个128b（64个uint8）
       uint32_t kv_idx = fz * 32 + tid % 4 * 2;
       v_smem.ldmatrix_m8n8x4(v_smem_offset_r, v_frag);
-      // layout
+      // 反量化 存储
       for (int i = 0; i < 4 / 2; i++) {
         T *v_tile_ptr0 = v_write_ptr + kv_idx * kv_t_stride + kv_head_idx * HEAD_DIM + dim_idx;
         T *v_tile_ptr1 = v_tile_ptr0 + 8;
-        convert_c8<T,IS_FP8>(frag_dq_T, v_frag[2 * i]); // 4 * uint8/fp8 -> 4 * T
-        convert_c8<T,IS_FP8>(frag_dq_T + 4, v_frag[2 * i + 1]); // 4 * uint8/fp8 -> 4 * T
         if (kv_idx < end_idx) {
+          convert_c8<T,IS_FP8>(frag_dq_T, v_frag[2 * i]); // 4个uint8/fp8 -> 4个T
+#ifdef C8_DEBUG
+          if (tid == 0 && wid == 0 && tile_idx == 0 && kv_head_idx == 0) {
+            printf("1.fy: %d, fz:%d, row_idx: %d, col_idx: %d, v_frag: %.f, %.f, %.f, %.f \n",
+              fy, fz, kv_idx, dim_idx, static_cast<float>(frag_dq_T[0]), static_cast<float>(frag_dq_T[1]),
+              static_cast<float>(frag_dq_T[2]), static_cast<float>(frag_dq_T[3]));
+          }
+#endif
           v_tile_ptr0[0] = frag_dq_T[0] * cache_v_scale;
-          v_tile_ptr1[0] = frag_dq_T[4] * cache_v_scale;
-        }
-        if (kv_idx + 1 < end_idx) {
           v_tile_ptr0[kv_t_stride] = frag_dq_T[1] * cache_v_scale;
-          v_tile_ptr1[kv_t_stride] = frag_dq_T[5] * cache_v_scale;
-        }
-        if (kv_idx + 8 < end_idx) {
           v_tile_ptr0[8 * kv_t_stride] = frag_dq_T[2] * cache_v_scale;
-          v_tile_ptr1[8 * kv_t_stride] = frag_dq_T[6] * cache_v_scale;
-        }
-        if (kv_idx + 9 < end_idx) {
           v_tile_ptr0[9 * kv_t_stride] = frag_dq_T[3] * cache_v_scale;
+
+          
+          convert_c8<T,IS_FP8>(frag_dq_T + 4, v_frag[2 * i + 1]); // 4个uint8/fp8 -> 4个T
+#ifdef C8_DEBUG
+          if (tid == 0 && wid == 0 && tile_idx == 0 && kv_head_idx == 0) {
+            printf("2.fy: %d, fz:%d, row_idx: %d, col_idx: %d, v_frag: %.f, %.f, %.f, %.f \n",
+              fy, fz, kv_idx, dim_idx + 8, static_cast<float>(frag_dq_T[4]), static_cast<float>(frag_dq_T[5]),
+              static_cast<float>(frag_dq_T[6]), static_cast<float>(frag_dq_T[7]));
+          }
+#endif
+          v_tile_ptr1[0] = frag_dq_T[4] * cache_v_scale;
+          v_tile_ptr1[kv_t_stride] = frag_dq_T[5] * cache_v_scale;
+          v_tile_ptr1[8 * kv_t_stride] = frag_dq_T[6] * cache_v_scale;
           v_tile_ptr1[9 * kv_t_stride] = frag_dq_T[7] * cache_v_scale;
         }
         kv_idx += 16;
@@ -520,250 +352,12 @@ __global__ void append_cache_kv_c8(
   }
 }
 
-template <typename T,
-          typename CacheT,
-          uint32_t HEAD_DIM,
-          uint32_t BLOCK_SIZE,
-          uint32_t NUM_WARPS=4>
-__global__ void append_cache_kv_c4(
-    const CacheT *__restrict__ cache_k,
-    const CacheT *__restrict__ cache_v,
-    T *__restrict__ k_out,
-    T *__restrict__ v_out,
-    const T *__restrict__ cache_k_dequant_scales,
-    const T *__restrict__ cache_v_dequant_scales,
-    const T *__restrict__ cache_k_zero_point,
-    const T *__restrict__ cache_v_zero_point,
-    const int *__restrict__ seq_lens_this_time,
-    const int *__restrict__ seq_lens_decoder,
-    const int *__restrict__ cu_seqlens_k,
-    const int *__restrict__ block_tables,
-    const int *batch_ids,
-    const int *tile_ids_per_batch,
-    const int max_blocks_per_seq,
-    const int kv_num_heads) {
-  // start_kv_idx: start kv_idx current block
-  // batch_id：block's batch_id
-  // TODO: 1.scale preload 2.frag_dq_T reuse 3.pipeline 4.store aligned 5.cacheT with template（int8/fp8)
-  const uint32_t tile_idx = blockIdx.x, kv_head_idx = blockIdx.z;
-  const uint32_t tid = threadIdx.x, wid = threadIdx.y;
-
-  const uint32_t batch_id = batch_ids[tile_idx];
-  const uint32_t start_kv_idx = tile_ids_per_batch[tile_idx] * BLOCK_SIZE;
-  const uint32_t end_idx = seq_lens_decoder[batch_id] - start_kv_idx;
-  if (seq_lens_this_time[batch_id] <= 0) {
-    return;
-  }
-
-  const int *cur_block_table = block_tables + batch_id * max_blocks_per_seq;
-  uint32_t block_id = cur_block_table[start_kv_idx / BLOCK_SIZE];
-  if (block_id < 0) block_id = 0;
-
-  constexpr uint32_t HEAD_DIM_HALF = HEAD_DIM / 2;
-  constexpr uint32_t BLOCK_SIZE_HALF = BLOCK_SIZE / 2;
-  // cache_kv idx
-  uint32_t kv_h_stride = BLOCK_SIZE * HEAD_DIM_HALF;
-  uint32_t block_stride = kv_num_heads * kv_h_stride;
-  const CacheT *cur_cache_k = cache_k + block_id * block_stride + kv_head_idx * kv_h_stride;
-  const CacheT *cur_cache_v = cache_v + block_id * block_stride + kv_head_idx * kv_h_stride;
-
-  // k_out v_out idx
-  uint32_t kv_t_stride = kv_num_heads * HEAD_DIM;
-  T *k_write_ptr = k_out + (cu_seqlens_k[batch_id] + start_kv_idx) * kv_t_stride;
-  T *v_write_ptr = v_out + (cu_seqlens_k[batch_id] + start_kv_idx) * kv_t_stride;
-
-  extern __shared__ uint8_t smem[];
-
-  uint32_t k_frag[4], v_frag[4], frag_dq[8];
-  T *frag_dq_T = reinterpret_cast<T *>(frag_dq);
-
-  // load dequant scales and zero points
-  const T *cache_k_scale_now = cache_k_dequant_scales + kv_head_idx * HEAD_DIM;
-  const T *cache_k_zp_now = cache_k_zero_point + kv_head_idx * HEAD_DIM;
-  const T *cache_v_scale_now = cache_v_dequant_scales + kv_head_idx * HEAD_DIM;
-  const T *cache_v_zp_now = cache_v_zero_point + kv_head_idx * HEAD_DIM;
-  T *cache_k_scale_smem = reinterpret_cast<T *>(
-      smem + BLOCK_SIZE * HEAD_DIM * sizeof(CacheT));
-  T *cache_k_zero_point_smem = cache_k_scale_smem + HEAD_DIM;
-  T *cache_v_scale_smem = cache_k_zero_point_smem + HEAD_DIM;
-  T *cache_v_zero_point_smem = cache_v_scale_smem + HEAD_DIM;
-#pragma unroll
-  for (uint32_t i = wid * 32 + tid; i < HEAD_DIM; i += 128) {
-    cache_k_scale_smem[i] = cache_k_scale_now[i];
-    cache_k_zero_point_smem[i] = cache_k_zp_now[i] + static_cast<T>(136.f);
-    cache_v_scale_smem[i] = cache_v_scale_now[i];
-    cache_v_zero_point_smem[i] = cache_v_zp_now[i] + static_cast<T>(136.f);
-  }
-
-  smem_t k_smem(smem);
-  constexpr uint32_t num_vecs_per_head_k =
-      HEAD_DIM_HALF / num_elems_per_128b<CacheT>(); // 2
-  constexpr uint32_t num_vecs_per_blocksize =
-      BLOCK_SIZE_HALF / num_elems_per_128b<CacheT>();
-  constexpr uint32_t inv_k_stride = 8 / num_vecs_per_head_k; // 4
-  constexpr uint32_t inv_v_stride = 8 / num_vecs_per_blocksize;
-
-  uint32_t k_smem_offset_w = smem_t::get_permuted_offset<num_vecs_per_head_k, inv_k_stride>(
-      wid * 8 + tid / 4, tid % 4);  // 2(iter) * 4(warp) * 8 row per warp
-
-  uint32_t k_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_head_k, inv_k_stride>(
-      wid * 16 + 8 * (tid / 16) + tid % 8, (tid % 16) / 8); //
-
-  uint32_t k_read_idx = (wid * 8 + tid / 4) * HEAD_DIM / 2 +
-                          tid % 4 * num_elems_per_128b<CacheT>();
-
-  // load k_smem 64 rows 128 cols
-  for (int fz = 0; fz < 2; fz++) { // 4 rows pre warp once, 16 rows all 4 warps once, need 4 iter
-    for (int fy = 0; fy < 1; fy++) { // 4 * 128b = 128 * int4 noce, need 1 iter
-      k_smem.load_128b_async<SharedMemFillMode::kNoFill>(
-            k_smem_offset_w, cur_cache_k + k_read_idx, end_idx > 0);
-      k_smem_offset_w =
-            k_smem.advance_offset_by_column<4, num_vecs_per_head_k>(k_smem_offset_w, fy);
-      k_read_idx += 4 * num_elems_per_128b<CacheT>();
-    }
-    k_smem_offset_w =
-          k_smem.advance_offset_by_row<8 * NUM_WARPS, num_vecs_per_head_k>(k_smem_offset_w) - 4;
-    k_read_idx += 8 * NUM_WARPS * HEAD_DIM / 2 - 4 * num_elems_per_128b<CacheT>();
-  }
-  commit_group();
-  wait_group<0>();
-  __syncthreads();
-
-  // deal k_smem 64 rows 128 cols
-  for (int fz = 0; fz < 1; fz++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 1 iter
-    uint32_t row_idx = wid * 16 + tid / 4;
-    for (int fy = 0; fy < 2; fy++) { // 2 * 128b = 64 * int4 noce, need 2 iter
-      uint32_t col_idx = fy * 64 + tid % 4 * 2;
-      k_smem.ldmatrix_m8n8x4(k_smem_offset_r, k_frag);
-
-
-      for (int i = 0; i < 2; i++) {
-        T *k_tile_ptr0 = k_write_ptr + row_idx * kv_t_stride + kv_head_idx * HEAD_DIM + col_idx;
-        T *k_tile_ptr1 = k_tile_ptr0 + 8 * kv_t_stride;
-        convert_int4(frag_dq_T, k_frag[2 * i]);
-        convert_int4(frag_dq_T + 8, k_frag[2 * i + 1]);
-
-        if (row_idx < end_idx) {
-          k_tile_ptr0[0] = (frag_dq_T[0] - cache_k_zero_point_smem[col_idx]) * cache_k_scale_smem[col_idx];
-          k_tile_ptr0[1] = (frag_dq_T[1] - cache_k_zero_point_smem[col_idx + 1]) * cache_k_scale_smem[col_idx + 1];
-          k_tile_ptr0[8] = (frag_dq_T[2] - cache_k_zero_point_smem[col_idx + 8]) * cache_k_scale_smem[col_idx + 8];
-          k_tile_ptr0[9] = (frag_dq_T[3] - cache_k_zero_point_smem[col_idx + 9]) * cache_k_scale_smem[col_idx + 9];
-          k_tile_ptr0[16] = (frag_dq_T[8] - cache_k_zero_point_smem[col_idx + 16]) * cache_k_scale_smem[col_idx + 16];
-          k_tile_ptr0[17] = (frag_dq_T[9] - cache_k_zero_point_smem[col_idx + 17]) * cache_k_scale_smem[col_idx + 17];
-          k_tile_ptr0[24] = (frag_dq_T[10] - cache_k_zero_point_smem[col_idx + 24]) * cache_k_scale_smem[col_idx + 24];
-          k_tile_ptr0[25] = (frag_dq_T[11] - cache_k_zero_point_smem[col_idx + 25]) * cache_k_scale_smem[col_idx + 25];
-        }
-
-        if (row_idx + 8 < end_idx) {
-          k_tile_ptr1[0] = (frag_dq_T[4] - cache_k_zero_point_smem[col_idx]) * cache_k_scale_smem[col_idx];
-          k_tile_ptr1[1] = (frag_dq_T[5] - cache_k_zero_point_smem[col_idx + 1]) * cache_k_scale_smem[col_idx + 1];
-          k_tile_ptr1[8] = (frag_dq_T[6] - cache_k_zero_point_smem[col_idx + 8]) * cache_k_scale_smem[col_idx + 8];
-          k_tile_ptr1[9] = (frag_dq_T[7] - cache_k_zero_point_smem[col_idx + 9]) * cache_k_scale_smem[col_idx + 9];
-          k_tile_ptr1[16] = (frag_dq_T[12] - cache_k_zero_point_smem[col_idx + 16]) * cache_k_scale_smem[col_idx + 16];
-          k_tile_ptr1[17] = (frag_dq_T[13] - cache_k_zero_point_smem[col_idx + 17]) * cache_k_scale_smem[col_idx + 17];
-          k_tile_ptr1[24] = (frag_dq_T[14] - cache_k_zero_point_smem[col_idx + 24]) * cache_k_scale_smem[col_idx + 24];
-          k_tile_ptr1[25] = (frag_dq_T[15] - cache_k_zero_point_smem[col_idx + 25]) * cache_k_scale_smem[col_idx + 25];
-        }
-        col_idx += 32;
-      }
-      k_smem_offset_r = k_smem.advance_offset_by_column<2, num_vecs_per_head_k>(
-        k_smem_offset_r, fy);
-    }
-    k_smem_offset_r =
-      k_smem.advance_offset_by_row<16 * NUM_WARPS, num_vecs_per_head_k>(k_smem_offset_r) - 4;
-  }
-
-  // ================v================
-  smem_t v_smem(smem + BLOCK_SIZE * HEAD_DIM * sizeof(CacheT) / 2);
-  uint32_t v_smem_offset_w = smem_t::get_permuted_offset<num_vecs_per_blocksize, inv_v_stride>(
-      wid * 16 + tid / 2, tid % 2);  // 4 * 8 per warp
-
-  uint32_t v_smem_offset_r = smem_t::get_permuted_offset<num_vecs_per_blocksize, inv_v_stride>(
-      wid * 16 + 8 * (tid / 16) + tid % 8, (tid % 16) / 8);
-
-  uint32_t v_read_idx = (wid * 16 + tid / 2) * BLOCK_SIZE_HALF +
-                          tid % 2 * num_elems_per_128b<CacheT>();
-  // load v_smem 128 rows 64 rows
-  for (int fy = 0; fy < 2; fy++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 2 iter
-    for (int fz = 0; fz < 1; fz++) { // 2 * 128b = 64 * int4 noce, need 1 iter
-      v_smem.load_128b_async<SharedMemFillMode::kNoFill>(
-              v_smem_offset_w, cur_cache_v + v_read_idx, end_idx > 0);
-      v_smem_offset_w =
-            v_smem.advance_offset_by_column<2, num_vecs_per_blocksize>(v_smem_offset_w, fz);
-      v_read_idx += 2 * num_elems_per_128b<CacheT>();
-    }
-    v_smem_offset_w =
-          v_smem.advance_offset_by_row<16 * NUM_WARPS, num_vecs_per_blocksize>(v_smem_offset_w) - 2;
-    v_read_idx += 16 * NUM_WARPS * BLOCK_SIZE_HALF - 2 * num_elems_per_128b<CacheT>();
-  }
-
-  commit_group();
-  wait_group<0>();
-  __syncthreads();
-
-  // deal v_smem 128 rows 64 cols
-  for (int fy = 0; fy < 2; fy++) { // 16 rows pre warp once, 64 rows all 4 warps once, need 2 iter
-    uint32_t dim_idx = fy * NUM_WARPS * 16 + wid * 16 + tid / 4;
-    for (int fz = 0; fz < 1; fz++) { // 2 * 128b = 64 * int4 noce, need 1 iter
-      uint32_t kv_idx = fz * 64 + tid % 4 * 2;
-      v_smem.ldmatrix_m8n8x4(v_smem_offset_r, v_frag);
-      // layout
-      for (int i = 0; i < 2; i++) {
-        T *v_tile_ptr0 = v_write_ptr + kv_idx * kv_t_stride + kv_head_idx * HEAD_DIM + dim_idx;
-        T *v_tile_ptr1 = v_tile_ptr0 + 8;
-
-        convert_int4(frag_dq_T, v_frag[2 * i]);
-        convert_int4(frag_dq_T + 8, v_frag[2 * i + 1]);
-        if (kv_idx < end_idx) {
-          v_tile_ptr0[0] = (frag_dq_T[0] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
-          v_tile_ptr1[0] = (frag_dq_T[4] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
-        }
-        if (kv_idx + 1 < end_idx) {
-          v_tile_ptr0[kv_t_stride] = (frag_dq_T[1] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
-          v_tile_ptr1[kv_t_stride] = (frag_dq_T[5] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
-        }
-        if (kv_idx + 8 < end_idx) {
-          v_tile_ptr0[8 * kv_t_stride] = (frag_dq_T[2] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
-          v_tile_ptr1[8 * kv_t_stride] = (frag_dq_T[6] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
-        }
-        if (kv_idx + 9 < end_idx) {
-          v_tile_ptr0[9 * kv_t_stride] = (frag_dq_T[3] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
-          v_tile_ptr1[9 * kv_t_stride] = (frag_dq_T[7] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
-        }
-        if (kv_idx + 16 < end_idx) {
-          v_tile_ptr0[16 * kv_t_stride] = (frag_dq_T[8] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
-          v_tile_ptr1[16 * kv_t_stride] = (frag_dq_T[12] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
-        }
-        if (kv_idx + 17 < end_idx) {
-          v_tile_ptr0[17 * kv_t_stride] = (frag_dq_T[9] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
-          v_tile_ptr1[17 * kv_t_stride] = (frag_dq_T[13] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
-        }
-        if (kv_idx + 24 < end_idx) {
-          v_tile_ptr0[24 * kv_t_stride] = (frag_dq_T[10] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
-          v_tile_ptr1[24 * kv_t_stride] = (frag_dq_T[14] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
-        }
-        if (kv_idx + 25 < end_idx) {
-          v_tile_ptr0[25 * kv_t_stride] = (frag_dq_T[11] - cache_v_zero_point_smem[dim_idx]) * cache_v_scale_smem[dim_idx];
-          v_tile_ptr1[25 * kv_t_stride] = (frag_dq_T[15] - cache_v_zero_point_smem[dim_idx + 8]) * cache_v_scale_smem[dim_idx + 8];
-        }
-        kv_idx += 32;
-      }
-      v_smem_offset_r = v_smem.advance_offset_by_column<2, num_vecs_per_blocksize>(
-          v_smem_offset_r, fz);
-    }
-    v_smem_offset_r =
-      v_smem.advance_offset_by_row<16 * NUM_WARPS, num_vecs_per_blocksize>(v_smem_offset_r) - 2;
-  }
-}
-
 template <typename T, uint32_t HEAD_DIM, uint32_t BLOCK_SIZE>
-void AppendCacheKV(
+void AppendDequantCache(
   const paddle::Tensor &cache_k,
   const paddle::Tensor &cache_v,
   const paddle::Tensor &cache_k_dequant_scales,
   const paddle::Tensor &cache_v_dequant_scales,
-  const paddle::Tensor &cache_k_zp,
-  const paddle::Tensor &cache_v_zp,
   const paddle::Tensor &seq_lens_this_time,
   const paddle::Tensor &seq_lens_decoder,
   const paddle::Tensor &cu_seqlens_k,
@@ -777,41 +371,19 @@ void AppendCacheKV(
   paddle::Tensor *k_out,
   paddle::Tensor *v_out,
   const cudaStream_t& stream
-) {
+) {  
   using NV_TYPE = typename cascade_attn_type_traits<T>::type;
-  constexpr int NUM_WARPS = 4;
-  int block_num = cache_num_blocks_x.data<int>()[0];
-  dim3 grids(block_num, 1, kv_num_heads);
-  dim3 blocks(32, NUM_WARPS);
-  if (cache_quant_type == "none") {
-    const uint32_t smem_size = BLOCK_SIZE * HEAD_DIM * sizeof(T) * 2;
-    auto kernel_func = append_cache_kv_c16<NV_TYPE, NV_TYPE, HEAD_DIM, BLOCK_SIZE, NUM_WARPS>;
-
-    if (smem_size >= 48 * 1024) {
-      cudaFuncSetAttribute(kernel_func,
-                          cudaFuncAttributeMaxDynamicSharedMemorySize,
-                          smem_size);
-    }
-    kernel_func<<<grids, blocks, smem_size, stream>>>(
-          reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_k.data<T>())),
-          reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_v.data<T>())),
-          reinterpret_cast<NV_TYPE *>(k_out->data<T>()),
-          reinterpret_cast<NV_TYPE *>(v_out->data<T>()),
-          seq_lens_this_time.data<int>(),
-          seq_lens_decoder.data<int>(),
-          cu_seqlens_k.data<int>(),
-          block_tables.data<int>(),
-          cache_batch_ids.data<int>(),
-          cache_tile_ids_per_batch.data<int>(),
-          max_blocks_per_seq,
-          kv_num_heads
-    );
-  } else if (cache_quant_type == "cache_int8" || cache_quant_type == "cache_fp8") {
+  if (cache_quant_type == "cache_int8" || cache_quant_type == "cache_fp8") {
+    constexpr int NUM_WARPS = 4;
+    int block_num = cache_num_blocks_x.data<int>()[0];
+    dim3 grids(block_num, 1, kv_num_heads);
+    dim3 blocks(32, NUM_WARPS);
+    
     const uint32_t smem_size = BLOCK_SIZE * HEAD_DIM * sizeof(uint8_t) * 2;
 
-    auto kernel_func = append_cache_kv_c8<NV_TYPE, uint8_t, HEAD_DIM, BLOCK_SIZE, NUM_WARPS, false>;
+    auto kernel_func = append_dequant_cache_kv_c8<NV_TYPE, uint8_t, HEAD_DIM, BLOCK_SIZE, NUM_WARPS, false>;
     if (cache_quant_type == "cache_fp8") {
-      kernel_func = append_cache_kv_c8<NV_TYPE, uint8_t, HEAD_DIM, BLOCK_SIZE, NUM_WARPS, true>;
+      kernel_func = append_dequant_cache_kv_c8<NV_TYPE, uint8_t, HEAD_DIM, BLOCK_SIZE, NUM_WARPS, true>;
     }
     if (smem_size >= 48 * 1024) {
       cudaFuncSetAttribute(kernel_func,
@@ -834,34 +406,6 @@ void AppendCacheKV(
           max_blocks_per_seq,
           kv_num_heads
     );
-  } else if (cache_quant_type == "cache_int4_zp") {
-    const uint32_t smem_size = BLOCK_SIZE * HEAD_DIM * sizeof(uint8_t) + 4 * HEAD_DIM * sizeof(T);
-
-    auto kernel_func = append_cache_kv_c4<NV_TYPE, uint8_t, HEAD_DIM, BLOCK_SIZE, NUM_WARPS>;
-
-    if (smem_size >= 48 * 1024) {
-      cudaFuncSetAttribute(kernel_func,
-                          cudaFuncAttributeMaxDynamicSharedMemorySize,
-                          smem_size);
-    }
-    kernel_func<<<grids, blocks, smem_size, stream>>>(
-          cache_k.data<uint8_t>(),
-          cache_v.data<uint8_t>(),
-          reinterpret_cast<NV_TYPE *>(k_out->data<T>()),
-          reinterpret_cast<NV_TYPE *>(v_out->data<T>()),
-          reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_k_dequant_scales.data<T>())),
-          reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_v_dequant_scales.data<T>())),
-          reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_k_zp.data<T>())),
-          reinterpret_cast<NV_TYPE *>(const_cast<T *>(cache_v_zp.data<T>())),
-          seq_lens_this_time.data<int>(),
-          seq_lens_decoder.data<int>(),
-          cu_seqlens_k.data<int>(),
-          block_tables.data<int>(),
-          cache_batch_ids.data<int>(),
-          cache_tile_ids_per_batch.data<int>(),
-          max_blocks_per_seq,
-          kv_num_heads
-    );
   } else {
     PADDLE_THROW("%s mode isn't implemented yet", cache_quant_type.c_str());
   }
@@ -877,7 +421,8 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
     const paddle::Tensor& seq_lens_this_time,
     const paddle::Tensor& seq_lens_encoder,
     const paddle::Tensor& seq_lens_decoder,
-    const paddle::Tensor& batch_id_per_token,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::Tensor& kv_batch_ids,
     const paddle::Tensor& kv_tile_ids,
@@ -894,8 +439,7 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
     const paddle::optional<paddle::Tensor>& kv_signal_data,
     const int kv_token_num,
     const int max_seq_len,
-    const std::string& cache_quant_type,
-    const bool rope_3d) {
+    const std::string& cache_quant_type) {
   typedef PDTraits<paddle::DataType::BFLOAT16> traits_;
   typedef typename traits_::DataType DataType_;
   typedef typename traits_::data_t data_t;
@@ -906,9 +450,9 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
   const int token_num = qkv_dims[0];
   const int max_blocks_per_seq = block_tables.dims()[1];
   const int block_size = key_cache.dims()[2];
-  const int batch_size = seq_lens_this_time.dims()[0];
+  const int batch_size = cum_offsets.dims()[0];
   const int kv_num_heads = key_cache_dims[1];
-  const int head_dim = cache_quant_type == "cache_int4_zp" ? key_cache_dims[3] * 2 : key_cache_dims[3];
+  const int head_dim = key_cache_dims[3];
   const int num_heads = qkv_dims[qkv_dims.size() - 1] / head_dim - 2 * kv_num_heads;
   const float softmax_scale = 1.f / sqrt(head_dim);
 
@@ -919,7 +463,7 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
   meta_data.q_num_heads = num_heads;
   meta_data.max_blocks_per_seq = max_blocks_per_seq;
   meta_data.block_size = block_size;
-  meta_data.batch_size = seq_lens_this_time.dims()[0];
+  meta_data.batch_size = cum_offsets.dims()[0];
 
   phi::GPUContext* dev_ctx = static_cast<phi::GPUContext*>(phi::DeviceContextPool::Instance().Get(qkv.place()));
 
@@ -949,58 +493,31 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
         v.data<data_t>(),
         qkv.data<data_t>(),
         rotary_embs.data<float>(),
-        batch_id_per_token.data<int>(),
+        padding_offsets.data<int>(),
         seq_lens_encoder.data<int>(),
         seq_lens_decoder.data<int>(),
-        cu_seqlens_q.data<int>(),
         cu_seqlens_k.data<int>(),
         token_num,
         num_heads,
         kv_num_heads,
         max_seq_len,
-        rope_3d ? rotary_embs.dims()[3] : rotary_embs.dims()[2],
+        rotary_embs.dims()[2],
         head_dim,
-        rope_3d,
         stream);
-
-  if (token_num < kv_token_num) {
-    AppendCacheKV<data_t, 128, 64>(
-      key_cache,
-      value_cache,
-      cache_k_dequant_scales.get(),
-      cache_v_dequant_scales.get(),
-      cache_k_zp.get(),
-      cache_v_zp.get(),
-      seq_lens_this_time,
-      seq_lens_decoder,
-      cu_seqlens_k,
-      block_tables,
-      cache_batch_ids,
-      cache_tile_ids,
-      cache_num_blocks,
-      max_blocks_per_seq,
-      kv_num_heads,
-      cache_quant_type,
-      &k,
-      &v,
-      stream
-    );
-  }
   // write cache
   if (cache_quant_type == "none") {
     CascadeAppendWriteCacheKVQKV<data_t>(
       meta_data,
       qkv_out,
       block_tables,
-      batch_id_per_token,
-      cu_seqlens_q,
+      padding_offsets,
       seq_lens_encoder,
       seq_lens_decoder,
       max_seq_len,
       stream,
       const_cast<paddle::Tensor*>(&key_cache),
       const_cast<paddle::Tensor*>(&value_cache));
-  } else if (cache_quant_type == "cache_int8" || cache_quant_type == "cache_fp8" || cache_quant_type == "block_wise_fp8") {
+  } else if (cache_quant_type == "cache_int8" || cache_quant_type == "cache_fp8") {
     CascadeAppendWriteCacheKVC8QKV<data_t, 128, 64>(
         meta_data,
         *const_cast<paddle::Tensor*>(&key_cache),
@@ -1010,44 +527,18 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
         cache_v_quant_scales.get(),
         seq_lens_this_time,
         seq_lens_decoder,
-        batch_id_per_token,
-        cu_seqlens_q,
+        padding_offsets,
+        cum_offsets,
         block_tables,
         kv_batch_ids,
         kv_tile_ids,
         kv_num_blocks_data,
         max_seq_len,
         false, // is_scale_channel_wise
-        cache_quant_type,
+        cache_quant_type == "cache_fp8", // is_fp8
         stream,
         const_cast<paddle::Tensor*>(&key_cache),
         const_cast<paddle::Tensor*>(&value_cache));
-  } else if (cache_quant_type == "cache_int4_zp") {
-    CascadeAppendWriteCacheKVC4QKV<data_t, 128, 64>(
-        meta_data,
-        *const_cast<paddle::Tensor*>(&key_cache),
-        *const_cast<paddle::Tensor*>(&value_cache),
-        qkv_out,
-        cache_k_quant_scales.get(),
-        cache_v_quant_scales.get(),
-        cache_k_zp.get(),
-        cache_v_zp.get(),
-        seq_lens_this_time,
-        seq_lens_decoder,
-        batch_id_per_token,
-        cu_seqlens_q,
-        block_tables,
-        kv_batch_ids,
-        kv_tile_ids,
-        kv_num_blocks_data,
-        max_seq_len,
-        stream,
-        const_cast<paddle::Tensor*>(&key_cache),
-        const_cast<paddle::Tensor*>(&value_cache));
-  } else {
-    PD_THROW(
-        "cache_quant_type_str should be one of [none, cache_int8, cache_fp8, "
-        "cache_int4_zp]");
   }
   const char* fmt_write_cache_completed_signal_str = std::getenv("FLAGS_fmt_write_cache_completed_signal");
   const char* FLAGS_use_pd_disaggregation_per_chunk = std::getenv("FLAGS_use_pd_disaggregation_per_chunk");
@@ -1068,6 +559,28 @@ std::vector<paddle::Tensor> GQARopeWriteCacheKernel(
         }
       }
   }
+
+  if (token_num < kv_token_num) {
+    AppendDequantCache<data_t, 128, 64>(
+      key_cache,
+      value_cache,
+      cache_k_dequant_scales.get(),
+      cache_v_dequant_scales.get(),
+      seq_lens_this_time,
+      seq_lens_decoder,
+      cu_seqlens_k,
+      block_tables,
+      cache_batch_ids,
+      cache_tile_ids,
+      cache_num_blocks,
+      max_blocks_per_seq,
+      kv_num_heads,
+      cache_quant_type,
+      &k,
+      &v,
+      stream
+    );
+  }
   return {q, k, v, qkv_out};
 }
 
@@ -1081,7 +594,8 @@ PD_BUILD_STATIC_OP(gqa_rope_write_cache)
              "seq_lens_this_time",
              "seq_lens_encoder",
              "seq_lens_decoder",
-             "batch_id_per_token",
+             "padding_offsets",
+             "cum_offsets",
              "block_tables",
              "kv_batch_ids",
              "kv_tile_ids_per_batch",

custom_ops/gpu_ops/append_attn/mla_cache_kernel.cu

@@ -1,292 +0,0 @@
-// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-#pragma once
-
-#include "helper.h"
-#include "mla_cache_kernel.cuh"
-
-template <paddle::DataType T>
-std::vector<paddle::Tensor> PrefillMLAWriteCache(
-                    const AppendAttnMetaData& meta_data,
-                    const paddle::Tensor& kv_nope,
-                    const paddle::Tensor& kv_pe,
-                    const paddle::Tensor& seq_lens,
-                    const paddle::Tensor& seq_lens_decoder,
-                    const paddle::Tensor& batch_id_per_token,
-                    const paddle::Tensor& cu_seqlens_q,
-                    const paddle::Tensor& block_tables,
-                    const int max_seq_len,
-                    cudaStream_t& stream,
-                    paddle::Tensor* kv_cache) {
-  typedef PDTraits<T> traits_;
-  typedef typename traits_::DataType DataType_;
-  typedef typename traits_::data_t data_t;
-
-  auto max_blocks_per_seq = meta_data.max_blocks_per_seq;
-  auto num_tokens = meta_data.token_nums;
-  auto block_size = meta_data.block_size;
-  auto nope_size = meta_data.head_dims_v;
-  auto all_size = meta_data.head_dims;
-  int pe_size = all_size - nope_size;
-  auto kv_num_heads = meta_data.kv_num_heads;
-  const uint32_t elem_nums = num_tokens * kv_num_heads * all_size;
-
-  constexpr int PackSize = 16 / sizeof(DataType_);
-  const int pack_num = elem_nums / PackSize;
-  const int blocksize = 128;
-  int grid_size = 1;
-  GetNumBlocks<128>(pack_num, &grid_size);
-
-  prefill_absorb_cache_kernel<DataType_, PackSize>
-      <<<grid_size, blocksize, 0, stream>>>(
-          reinterpret_cast<DataType_*>(const_cast<data_t*>(kv_nope.data<data_t>())),
-          reinterpret_cast<DataType_*>(const_cast<data_t*>(kv_pe.data<data_t>())),
-          reinterpret_cast<DataType_*>(kv_cache->data<data_t>()),
-          block_tables.data<int>(),
-          batch_id_per_token.data<int>(),
-          cu_seqlens_q.data<int>(),
-          seq_lens.data<int>(),
-          seq_lens_decoder.data<int>(),
-          max_seq_len,
-          max_blocks_per_seq,
-          kv_num_heads,
-          nope_size,
-          pe_size,
-          block_size,
-          elem_nums);
-  return {};
-}
-
-std::vector<paddle::Tensor> PrefillMLAWriteCacheKernel(
-    const paddle::Tensor& kv_nope,
-    const paddle::Tensor& kv_pe,
-    const paddle::Tensor& kv_cache,
-    const paddle::Tensor& seq_lens,
-    const paddle::Tensor& seq_lens_decoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
-    const paddle::Tensor& block_tables,
-    const std::string& cache_quant_type_str,
-    const int max_seq_len) {
-  cudaStream_t stream = kv_pe.stream();
-  AppendAttnMetaData meta_data;
-  const auto& kv_nope_dims = kv_nope.dims();
-  const auto& kv_pe_dims = kv_pe.dims();
-  const auto& kv_cache_dims = kv_cache.dims();
-  meta_data.kv_num_heads = kv_cache_dims[1];
-  const auto nope_size = kv_nope_dims[kv_nope_dims.size() - 1] / meta_data.kv_num_heads;
-  meta_data.token_nums = kv_nope_dims[0];
-  meta_data.head_dims = kv_cache_dims[3];
-  meta_data.head_dims_v = nope_size;
-
-  meta_data.max_blocks_per_seq = block_tables.dims()[1];
-  meta_data.block_size = kv_cache_dims[2];
-  meta_data.batch_size = seq_lens_decoder.dims()[0];
-  switch (kv_pe.dtype()) {
-    case paddle::DataType::BFLOAT16: {
-      return PrefillMLAWriteCache<paddle::DataType::BFLOAT16>(meta_data,
-                              kv_nope,
-                              kv_pe,
-                              seq_lens,
-                              seq_lens_decoder,
-                              batch_id_per_token,
-                              cu_seqlens_q,
-                              block_tables,
-                              max_seq_len,
-                              stream,
-                              const_cast<paddle::Tensor*>(&kv_cache));
-    }
-    case paddle::DataType::FLOAT16: {
-      return PrefillMLAWriteCache<paddle::DataType::FLOAT16>(meta_data,
-                              kv_nope,
-                              kv_pe,
-                              seq_lens,
-                              seq_lens_decoder,
-                              batch_id_per_token,
-                              cu_seqlens_q,
-                              block_tables,
-                              max_seq_len,
-                              stream,
-                              const_cast<paddle::Tensor*>(&kv_cache));
-    }
-  }
-  return {};
-}
-
-template <paddle::DataType T>
-std::vector<paddle::Tensor> DecodeMLAWriteCache(
-                    const AppendAttnMetaData& meta_data,
-                    const paddle::Tensor& kv_nope,
-                    const paddle::Tensor& kv_pe,
-                    const paddle::Tensor& seq_lens,
-                    const paddle::Tensor& seq_lens_encoder,
-                    const paddle::Tensor& batch_id_per_token,
-                    const paddle::Tensor& cu_seqlens_q,
-                    const paddle::Tensor& block_tables,
-                    const int max_seq_len,
-                    const bool speculate_decoder,
-                    cudaStream_t& stream,
-                    paddle::Tensor* kv_cache) {
-  typedef PDTraits<T> traits_;
-  typedef typename traits_::DataType DataType_;
-  typedef typename traits_::data_t data_t;
-
-  auto max_blocks_per_seq = meta_data.max_blocks_per_seq;
-  auto bsz = meta_data.batch_size;
-  auto token_num = meta_data.token_nums;
-  auto block_size = meta_data.block_size;
-  auto nope_size = meta_data.head_dims_v;
-  auto all_size = meta_data.head_dims;
-  int pe_size = all_size - nope_size;
-  auto kv_num_heads = meta_data.kv_num_heads;
-  constexpr int PackSize = 16 / sizeof(DataType_);
-  const int blocksize = 128;
-  int grid_size = 1;
-
-
-  if (speculate_decoder) {
-    const uint32_t elem_nums = token_num * kv_num_heads * all_size;
-    const int pack_num = elem_nums / PackSize;
-    GetNumBlocks<128>(pack_num, &grid_size);
-    speculate_decode_absorb_cache_kernel<DataType_, PackSize>
-        <<<grid_size, blocksize, 0, stream>>>(
-            reinterpret_cast<DataType_*>(const_cast<data_t*>(kv_nope.data<data_t>())),
-            reinterpret_cast<DataType_*>(const_cast<data_t*>(kv_pe.data<data_t>())),
-            reinterpret_cast<DataType_*>(kv_cache->data<data_t>()),
-            block_tables.data<int>(),
-            batch_id_per_token.data<int>(),
-            cu_seqlens_q.data<int>(),
-            seq_lens.data<int>(),
-            seq_lens_encoder.data<int>(),
-            max_seq_len,
-            max_blocks_per_seq,
-            kv_num_heads,
-            nope_size,
-            pe_size,
-            block_size,
-            elem_nums);
-  } else {
-    const uint32_t elem_nums = bsz * kv_num_heads * all_size;
-    const int pack_num = elem_nums / PackSize;
-    GetNumBlocks<128>(pack_num, &grid_size);
-    decode_absorb_cache_kernel<DataType_, PackSize>
-        <<<grid_size, blocksize, 0, stream>>>(
-            reinterpret_cast<DataType_*>(const_cast<data_t*>(kv_nope.data<data_t>())),
-            reinterpret_cast<DataType_*>(const_cast<data_t*>(kv_pe.data<data_t>())),
-            reinterpret_cast<DataType_*>(kv_cache->data<data_t>()),
-            block_tables.data<int>(),
-            cu_seqlens_q.data<int>(),
-            seq_lens.data<int>(),
-            seq_lens_encoder.data<int>(),
-            max_seq_len,
-            max_blocks_per_seq,
-            kv_num_heads,
-            nope_size,
-            pe_size,
-            block_size,
-            elem_nums);
-  }
-  return {};
-}
-
-std::vector<paddle::Tensor> DecodeMLAWriteCacheKernel(
-    const paddle::Tensor& kv_nope,
-    const paddle::Tensor& kv_pe,
-    const paddle::Tensor& kv_cache,
-    const paddle::Tensor& seq_lens,
-    const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
-    const paddle::Tensor& block_tables,
-    const std::string& cache_quant_type_str,
-    const int max_seq_len,
-    const bool speculate_decoder) {
-  cudaStream_t stream = kv_pe.stream();
-  AppendAttnMetaData meta_data;
-  const auto& kv_nope_dims = kv_nope.dims();
-  const auto& kv_pe_dims = kv_pe.dims();
-  const auto& kv_cache_dims = kv_cache.dims();
-  meta_data.kv_num_heads = kv_cache_dims[1];
-  const auto nope_size = kv_nope_dims[kv_nope_dims.size() - 1] / meta_data.kv_num_heads;
-  meta_data.token_nums = kv_nope_dims[0];
-  meta_data.head_dims = kv_cache_dims[3];
-  meta_data.head_dims_v = nope_size;
-
-  meta_data.max_blocks_per_seq = block_tables.dims()[1];
-  meta_data.block_size = kv_cache_dims[2];
-  meta_data.batch_size = seq_lens_encoder.dims()[0];
-  switch (kv_pe.dtype()) {
-    case paddle::DataType::BFLOAT16: {
-      return DecodeMLAWriteCache<paddle::DataType::BFLOAT16>(meta_data,
-                              kv_nope,
-                              kv_pe,
-                              seq_lens,
-                              seq_lens_encoder,
-                              batch_id_per_token,
-                              cu_seqlens_q,
-                              block_tables,
-                              max_seq_len,
-                              speculate_decoder,
-                              stream,
-                              const_cast<paddle::Tensor*>(&kv_cache));
-    }
-    case paddle::DataType::FLOAT16: {
-      return DecodeMLAWriteCache<paddle::DataType::FLOAT16>(meta_data,
-                              kv_nope,
-                              kv_pe,
-                              seq_lens,
-                              seq_lens_encoder,
-                              batch_id_per_token,
-                              cu_seqlens_q,
-                              block_tables,
-                              max_seq_len,
-                              speculate_decoder,
-                              stream,
-                              const_cast<paddle::Tensor*>(&kv_cache));
-    }
-  }
-  return {};
-}
-
-
-PD_BUILD_STATIC_OP(prefill_mla_write_cache)
-    .Inputs({"kv_nope",
-             "kv_pe",
-             "kv_cache",
-             "seq_lens",
-             "seq_lens_decoder",
-             "batch_id_per_token",
-             "cu_seqlens_q",
-             "block_tables"})
-    .Outputs({"kv_cache_out"})
-    .SetInplaceMap({{"kv_cache", "kv_cache_out"}})
-    .Attrs({"cache_quant_type_str: std::string",
-            "max_seq_len: int"})
-    .SetKernelFn(PD_KERNEL(PrefillMLAWriteCacheKernel));
-
-PD_BUILD_STATIC_OP(decode_mla_write_cache)
-    .Inputs({"kv_nope",
-             "kv_pe",
-             "kv_cache",
-             "seq_lens",
-             "seq_lens_encoder",
-             "batch_id_per_token",
-             "cu_seqlens_q",
-             "block_tables"})
-    .Outputs({"kv_cache_out"})
-    .SetInplaceMap({{"kv_cache", "kv_cache_out"}})
-    .Attrs({"cache_quant_type_str: std::string",
-            "max_seq_len: int",
-            "speculate_decoder: bool"})
-    .SetKernelFn(PD_KERNEL(DecodeMLAWriteCacheKernel));

custom_ops/gpu_ops/append_attn/mla_cache_kernel.cuh

@@ -1,240 +0,0 @@
-// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-#pragma once
-
-#include "helper.h"
-#include "mem_util.cuh"
-#include "utils.cuh"
-
-template <typename T, int VecSize = 1>
-__global__ void decode_absorb_cache_kernel(
-    const T* __restrict__ kv_nope,  // [bsz, kv_num_heads, pe_size] 512
-    const T* __restrict__ kv_pe,  // [bsz, kv_num_heads, nope_size] 64
-    T* __restrict__ kv_cache,    // [num_blocks, kv_num_heads, block_size,
-                                  // nope_size]
-    const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ cu_seqlens_q,
-    const int* __restrict__ seq_lens,          // [bsz]
-    const int* __restrict__ seq_lens_encoder,  // [bsz]
-    const int max_seq_len,
-    const int max_blocks_per_seq,
-    const int kv_num_heads,
-    const int nope_size,
-    const int pe_size,
-    const int block_size,
-    const uint32_t elem_cnt) {
-  using LoadT = AlignedVector<T, VecSize>;
-  constexpr int HalfVecSize = VecSize / 2;
-  LoadT src_vec;
-
-  int64_t global_thread_idx = blockDim.x * blockIdx.x + threadIdx.x;
-  const uint32_t nope_hidden_size = kv_num_heads * nope_size;
-  const uint32_t pe_hidden_size = kv_num_heads * pe_size;
-  const uint32_t all_size = nope_size + pe_size;
-  const int64_t hidden_size = nope_hidden_size + pe_hidden_size;
-
-  for (int32_t linear_index = global_thread_idx * VecSize,
-               step = gridDim.x * blockDim.x * VecSize;
-       linear_index < elem_cnt;
-       linear_index += step) {
-    const int ori_bi = linear_index / hidden_size;
-    const int bias = linear_index % hidden_size;
-    const int start_token_idx = cu_seqlens_q[ori_bi];
-    if (seq_lens_encoder[ori_bi] > 0) return;
-    const int write_seq_id = seq_lens[ori_bi];
-
-    if (write_seq_id == 0) continue;
-
-    const int* block_table_now = nullptr;
-
-    block_table_now = block_tables + ori_bi * max_blocks_per_seq;
-    const int block_idx = block_table_now[write_seq_id / block_size];
-    const int block_offset = write_seq_id % block_size;
-
-    if (bias < nope_hidden_size) { // pe
-      const uint32_t inner_bias = bias;
-      const uint32_t hi = inner_bias / nope_size;
-      const uint32_t h_bias = inner_bias % nope_size;
-      const uint32_t tgt_idx = block_idx * kv_num_heads * block_size * all_size +
-                             hi * block_size * all_size +
-                             block_offset * all_size + h_bias;
-      const uint32_t ori_idx =
-          start_token_idx * nope_hidden_size + inner_bias;
-      Load<T, VecSize>(&kv_nope[ori_idx], &src_vec);
-      Store<T, VecSize>(src_vec, &kv_cache[tgt_idx]);
-    } else {
-      const uint32_t inner_bias = bias - nope_hidden_size;
-      const uint32_t hi = inner_bias / pe_size;
-      const uint32_t h_bias = inner_bias % pe_size;
-      const uint32_t tgt_idx = block_idx * kv_num_heads * block_size * all_size +
-                             hi * block_size * all_size +
-                             block_offset * all_size + nope_size + h_bias;
-      const uint32_t ori_idx =
-          start_token_idx * pe_hidden_size + inner_bias;
-      Load<T, VecSize>(&kv_pe[ori_idx], &src_vec);
-      Store<T, VecSize>(src_vec, &kv_cache[tgt_idx]);
-    }
-  }
-}
-
-template <typename T, int VecSize = 1>
-__global__ void speculate_decode_absorb_cache_kernel(
-    const T* __restrict__ kv_nope,  // [bsz, kv_num_heads, pe_size] 512
-    const T* __restrict__ kv_pe,  // [bsz, kv_num_heads, nope_size] 64
-    T* __restrict__ kv_cache,    // [num_blocks, kv_num_heads, block_size,
-                                  // nope_size]
-    const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,
-    const int* __restrict__ cu_seqlens_q,
-    const int* __restrict__ seq_lens,          // [bsz]
-    const int* __restrict__ seq_lens_encoder,  // [bsz]
-    const int max_seq_len,
-    const int max_blocks_per_seq,
-    const int kv_num_heads,
-    const int nope_size,
-    const int pe_size,
-    const int block_size,
-    const uint32_t elem_cnt) {
-  using LoadT = AlignedVector<T, VecSize>;
-  constexpr int HalfVecSize = VecSize / 2;
-  LoadT src_vec;
-
-  int64_t global_thread_idx = blockDim.x * blockIdx.x + threadIdx.x;
-  const uint32_t nope_hidden_size = kv_num_heads * nope_size;
-  const uint32_t pe_hidden_size = kv_num_heads * pe_size;
-  const uint32_t all_size = nope_size + pe_size;
-  const int64_t hidden_size = nope_hidden_size + pe_hidden_size;
-
-  for (int32_t linear_index = global_thread_idx * VecSize,
-               step = gridDim.x * blockDim.x * VecSize;
-       linear_index < elem_cnt;
-       linear_index += step) {
-    const int token_id = linear_index / hidden_size;
-    const int ori_bi = batch_id_per_token[token_id];
-    if (seq_lens[ori_bi] == 0) continue;
-    const int bias = linear_index % hidden_size;
-    const int start_token_idx = cu_seqlens_q[ori_bi];
-    const int write_seq_id =
-        seq_lens[ori_bi] + token_id - start_token_idx;
-    if (write_seq_id == 0) continue;
-
-    const int* block_table_now = nullptr;
-
-    block_table_now = block_tables + ori_bi * max_blocks_per_seq;
-    const int block_idx = block_table_now[write_seq_id / block_size];
-    const int block_offset = write_seq_id % block_size;
-    if (block_idx < 0) {
-      printf(
-          "Fatal Error!!!, block idx %d when write_seq_id is %d\n some key var "
-          "%d %d %d %d\n",
-          block_idx,
-          write_seq_id,
-          ori_bi,
-          seq_lens[ori_bi],
-          token_id,
-          cu_seqlens_q[ori_bi]);
-    }
-    if (bias < nope_hidden_size) { // pe
-      const uint32_t inner_bias = bias;
-      const uint32_t hi = inner_bias / nope_size;
-      const uint32_t h_bias = inner_bias % nope_size;
-      const uint32_t tgt_idx = block_idx * kv_num_heads * block_size * all_size +
-                             hi * block_size * all_size +
-                             block_offset * all_size + h_bias;
-      const uint32_t ori_idx =
-          token_id * nope_hidden_size + inner_bias;
-      Load<T, VecSize>(&kv_nope[ori_idx], &src_vec);
-      Store<T, VecSize>(src_vec, &kv_cache[tgt_idx]);
-    } else {
-      const uint32_t inner_bias = bias - nope_hidden_size;
-      const uint32_t hi = inner_bias / pe_size;
-      const uint32_t h_bias = inner_bias % pe_size;
-      const uint32_t tgt_idx = block_idx * kv_num_heads * block_size * all_size +
-                             hi * block_size * all_size +
-                             block_offset * all_size + nope_size + h_bias;
-      const uint32_t ori_idx =
-          token_id * pe_hidden_size + inner_bias;
-      Load<T, VecSize>(&kv_pe[ori_idx], &src_vec);
-      Store<T, VecSize>(src_vec, &kv_cache[tgt_idx]);
-    }
-  }
-}
-
-template <typename T, int VecSize = 1>
-__global__ void prefill_absorb_cache_kernel(
-    const T* __restrict__ kv_nope,  // [bsz, kv_num_heads, pe_size] 512
-    const T* __restrict__ kv_pe,  // [bsz, kv_num_heads, nope_size] 64
-    T* __restrict__ kv_cache,    // [num_blocks, kv_num_heads, block_size,
-                                  // nope_size]
-    const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,
-    const int* __restrict__ cu_seqlens_q,
-    const int* __restrict__ seq_lens,          // [bsz]
-    const int* __restrict__ seq_lens_decoder,  // [bsz]
-    const int max_seq_len,
-    const int max_blocks_per_seq,
-    const int kv_num_heads,
-    const int nope_size,
-    const int pe_size,
-    const int block_size,
-    const uint32_t elem_cnt) {
-  using LoadT = AlignedVector<T, VecSize>;
-  LoadT src_vec;
-
-  int64_t global_thread_idx = blockDim.x * blockIdx.x + threadIdx.x;
-  const uint32_t nope_hidden_size = kv_num_heads * nope_size;
-  const uint32_t pe_hidden_size = kv_num_heads * pe_size;
-  const uint32_t all_size = nope_size + pe_size;
-  const int64_t hidden_size = nope_hidden_size + pe_hidden_size;
-
-  for (int32_t linear_index = global_thread_idx * VecSize,
-               step = gridDim.x * blockDim.x * VecSize;
-       linear_index < elem_cnt;
-       linear_index += step) {
-    const uint32_t token_idx = linear_index / hidden_size;
-    const uint32_t bias = linear_index % hidden_size;
-    const uint32_t ori_bi = batch_id_per_token[token_idx];
-    if (seq_lens[ori_bi] == 0) continue;
-    const uint32_t ori_seq_id = (token_idx - cu_seqlens_q[ori_bi]) + seq_lens_decoder[ori_bi];
-
-    const int* block_table_now = nullptr;
-    block_table_now = block_tables + ori_bi * max_blocks_per_seq;
-    const uint32_t block_idx = block_table_now[ori_seq_id / block_size];
-    const uint32_t block_offset = ori_seq_id % block_size;
-
-    if (bias < nope_hidden_size) { // pe
-      const uint32_t inner_bias = bias;
-      const uint32_t hi = inner_bias / nope_size;
-      const uint32_t h_bias = inner_bias % nope_size;
-      const uint32_t tgt_idx = block_idx * kv_num_heads * block_size * all_size +
-                             hi * block_size * all_size +
-                             block_offset * all_size + h_bias;
-      const uint32_t ori_idx =
-          token_idx * nope_hidden_size + inner_bias;
-      Load<T, VecSize>(&kv_nope[ori_idx], &src_vec);
-      Store<T, VecSize>(src_vec, &kv_cache[tgt_idx]);
-    } else {
-      const uint32_t inner_bias = bias - nope_hidden_size;
-      const uint32_t hi = inner_bias / pe_size;
-      const uint32_t h_bias = inner_bias % pe_size;
-      const uint32_t tgt_idx = block_idx * kv_num_heads * block_size * all_size +
-                             hi * block_size * all_size +
-                             block_offset * all_size + nope_size + h_bias;
-      const uint32_t ori_idx =
-          token_idx * pe_hidden_size + inner_bias;
-      Load<T, VecSize>(&kv_pe[ori_idx], &src_vec);
-      Store<T, VecSize>(src_vec, &kv_cache[tgt_idx]);
-    }
-  }
-}

custom_ops/gpu_ops/append_attn/multi_head_latent_attention_kernel.h

@@ -1,38 +0,0 @@
-// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-#pragma once
-#include "helper.h"
-#include "utils.cuh"
-
-template <typename T>
-void DecodeMLAAttentionKernel(
-    const AppendAttnMetaData& meta_data,
-    const paddle::Tensor &q, // [token_num, num_heads, head_dim]
-    const paddle::Tensor &cache_k,
-    const paddle::Tensor &cache_v,
-    const paddle::optional<paddle::Tensor>& attn_mask,
-    const paddle::optional<paddle::Tensor>& shift_bias,
-    const paddle::optional<paddle::Tensor>& smooth_weight,
-    const paddle::Tensor &seq_lens_q, // q_seq_len is 1
-    const paddle::Tensor &seq_lens_kv,
-    const paddle::Tensor &batch_id_per_token,
-    const paddle::Tensor &cu_seqlens_q,
-    const paddle::Tensor &block_table,
-    int max_seq_len,
-    int max_dec_len,
-    float softmax_scale,
-    float in_scale,
-    bool causal,
-    cudaStream_t &stream,
-    paddle::Tensor *out);

custom_ops/gpu_ops/append_attn/speculate_write_cache_with_rope_impl.cuh

@@ -18,168 +18,6 @@
 #include "mma_tensor_op.cuh"
 #include "utils.cuh"
 
-template <typename T, int VecSize = 1, typename InT = T>
-__global__ void append_speculate_cache_T_rope_qk_norm_kernel(
-    const InT* __restrict__ qkv,  // [token_num, num_heads + 2 * gqa_group_size,
-                                  // head_size]
-    T* __restrict__ key_cache,    // [num_blocks, gqa_group_size, block_size,
-                                  // head_size // 2]
-    T* __restrict__ value_cache,  // [num_blocks, gqa_group_size, block_size,
-                                  // head_size // 2]
-    T* __restrict__ q_out,
-    const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
-    const int* __restrict__ seq_lens_decoder,  // [bsz]
-    const float* __restrict__ cos_emb,
-    const float* __restrict__ sin_emb,
-    const float*
-        qkv_out_scales,   // [(num_heads + 2 * gqa_group_size) * head_size]
-    const T* qkv_biases,  // [num_head + 2 * gqa_group_size, dim_head]
-    const int max_seq_len,
-    const int max_blocks_per_seq,
-    const int num_heads,
-    const int output_inner_dim,
-    const int head_size,
-    const int block_size,
-    const int elem_cnt,
-    const int gqa_group_size,
-    const float* q_norm_weight,
-    const float* k_norm_weight,
-    const float rms_norm_eps,
-    const bool rope_3d) {
-  using LoadT = AlignedVector<T, VecSize>;
-  using LoadFloat = AlignedVector<float, VecSize>;
-  using LoadInT = AlignedVector<InT, VecSize>;
-  constexpr int HalfVecSize = VecSize / 2;
-  using LoadEmbT = AlignedVector<float, HalfVecSize>;
-  LoadInT src_vec;
-  LoadFloat scale_vec;
-  LoadT bias_vec;
-  LoadEmbT cos_emb_vec;
-  LoadEmbT sin_emb_vec;
-  LoadFloat tmp_vec;
-  LoadFloat q_norm_vec;
-  LoadFloat k_norm_vec;
-
-  int64_t global_warp_idx = blockDim.y * blockIdx.x + threadIdx.y;
-  int64_t all_warp_num = gridDim.x * blockDim.y;
-  int64_t all_head_dim = elem_cnt / head_size;
-
-  const int64_t hidden_size = (num_heads + 2 * gqa_group_size) * head_size;
-  const int half_head_size = head_size / 2;
-  for (int global_hi = global_warp_idx; global_hi < all_head_dim; global_hi += all_warp_num) {
-    int64_t linear_index = global_hi * head_size + threadIdx.x * VecSize;
-    const int token_id = linear_index / hidden_size;
-    const int ori_bi = batch_id_per_token[token_id];
-    if (seq_lens_decoder[ori_bi] == 0) continue;
-    const int bias = linear_index % hidden_size;
-    const int hi = bias / head_size;  // q + k + v
-    const int h_bias = bias % head_size;
-    const int start_token_idx = cu_seqlens_q[ori_bi];
-    const int write_seq_id =
-        seq_lens_decoder[ori_bi] + token_id - start_token_idx;
-    if (write_seq_id == 0) continue;
-
-    const int* block_table_now = block_tables + ori_bi * max_blocks_per_seq;
-    const int block_idx = block_table_now[write_seq_id / block_size];
-    if (block_idx < 0) {
-      printf(
-          "Fatal Error!!!, block idx %d when write_seq_id is %d\n some key var "
-          "%d %d %d %d\n",
-          block_idx,
-          write_seq_id,
-          ori_bi,
-          seq_lens_decoder[ori_bi],
-          token_id,
-          cu_seqlens_q[ori_bi]);
-    }
-    const int block_offset = write_seq_id % block_size;
-
-    const int write_q_idx =
-        token_id * output_inner_dim * head_size + hi * head_size + h_bias;
-
-    const int bias_idx = hi * head_size + h_bias;
-    Load<InT, VecSize>(&qkv[linear_index], &src_vec);
-    if (qkv_biases) {
-      Load<T, VecSize>(&qkv_biases[bias_idx], &bias_vec);
-    }
-    if (qkv_out_scales) {
-      Load<float, VecSize>(&qkv_out_scales[bias_idx], &scale_vec);
-    }
-    if (hi < num_heads + gqa_group_size) {
-      // q k rope
-      const int64_t emb_idx = write_seq_id * half_head_size + h_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + ori_bi * max_seq_len * head_size : emb_idx;
-      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
-    }
-    float thread_m2 = 0.0f;
-    float warp_m2 = 0.0f;
-#pragma unroll
-    for (int i = 0; i < HalfVecSize; i++) {
-      // add_bias + rope
-      float input_left = static_cast<float>(src_vec[2 * i]);
-      float input_right = static_cast<float>(src_vec[2 * i + 1]);
-      if (qkv_out_scales) {
-        input_left *= scale_vec[2 * i];
-        input_right *= scale_vec[2 * i + 1];
-      }
-      if (qkv_biases) {
-        input_left = input_left + static_cast<float>(bias_vec[2 * i]);
-        input_right = input_right + static_cast<float>(bias_vec[2 * i + 1]);
-      }
-      if (hi < num_heads + gqa_group_size) {
-        const float cos_tmp = cos_emb_vec[i];
-        const float sin_tmp = sin_emb_vec[i];
-        float tmp1 = input_left * cos_tmp - input_right * sin_tmp;
-        float tmp2 = input_right * cos_tmp + input_left * sin_tmp;
-        thread_m2 += tmp1 * tmp1 + tmp2 * tmp2;
-        tmp_vec[2 * i] = tmp1;
-        tmp_vec[2 * i + 1] = tmp2;
-      } else {
-        bias_vec[2 * i] = static_cast<T>(input_left);
-        bias_vec[2 * i + 1] = static_cast<T>(input_right);
-      }
-    }
-    if (hi < (num_heads + gqa_group_size)) {
-      WelfordWarpAllReduce<float, 32>(thread_m2, &warp_m2);
-      float row_variance =
-          max(warp_m2 / head_size, 0.0f);
-      float row_inv_var = Rsqrt(row_variance + rms_norm_eps);
-      if (hi < num_heads) {
-        Load<float, VecSize>(&q_norm_weight[threadIdx.x * VecSize], &q_norm_vec);
-        #pragma unroll
-        for (int i = 0; i < VecSize; i++) {
-          bias_vec[i] = static_cast<T>(tmp_vec[i] * row_inv_var * q_norm_vec[i]);
-        }
-      } else {
-        Load<float, VecSize>(&k_norm_weight[threadIdx.x * VecSize], &k_norm_vec);
-        #pragma unroll
-        for (int i = 0; i < VecSize; i++) {
-          bias_vec[i] = static_cast<T>(tmp_vec[i] * row_inv_var * k_norm_vec[i]);
-        }
-      }
-    }
-    if (hi < num_heads) {
-      // write q
-      Store<T, VecSize>(bias_vec, &q_out[write_q_idx]);
-    } else {
-      //  write k/v
-      const int kv_head_idx = (hi - num_heads) % gqa_group_size;
-      const int tgt_idx = (block_idx * gqa_group_size * block_size * head_size +
-                           kv_head_idx * block_size * head_size +
-                           block_offset * head_size + h_bias);
-      // write
-      if (hi < num_heads + gqa_group_size) {
-        Store<T, VecSize>(bias_vec, &key_cache[tgt_idx]);
-      } else {
-        Store<T, VecSize>(bias_vec, &value_cache[tgt_idx]);
-      }
-    }
-  }
-}
-
 template <int VecSize = 4, int HeadDim = 128>
 __global__ void append_clear_cache_int8_block(
     uint8_t* __restrict__ key_cache,    // [num_blocks, gqa_group_size,
@@ -188,8 +26,8 @@ __global__ void append_clear_cache_int8_block(
                                         // block_size, head_size // 2]
     const int* __restrict__ seq_lens,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const int max_seq_len,
     const int max_blocks_per_seq,
@@ -203,10 +41,10 @@ __global__ void append_clear_cache_int8_block(
   const int wid = tid / 32;
   const int lane_id = tid % 32;
   const int token_id = blockIdx.x;
+  const int ori_token_id = token_id + padding_offsets[token_id];
+  const int bid = ori_token_id / max_seq_len;
 
-  const int bid = batch_id_per_token[token_id];
-
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - cum_offsets[bid];
   const int head_idx = blockIdx.y * NUM_WARPS + wid;
 
   if (seq_lens_encoder[bid] > 0) return;
@@ -262,8 +100,8 @@ __global__ void append_clear_cache_int4_block(
                                         // block_size, head_size // 2]
     const int* __restrict__ seq_lens,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const int max_seq_len,
     const int max_blocks_per_seq,
@@ -277,10 +115,10 @@ __global__ void append_clear_cache_int4_block(
   const int wid = tid / 32;
   const int lane_id = tid % 32;
   const int token_id = blockIdx.x;
+  const int ori_token_id = token_id + padding_offsets[token_id];
+  const int bid = ori_token_id / max_seq_len;
 
-  const int bid = batch_id_per_token[token_id];
-
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - cum_offsets[bid];
   const int head_idx = blockIdx.y * NUM_WARPS + wid;
 
   if (seq_lens_encoder[bid] > 0) return;
@@ -340,8 +178,8 @@ __global__ void append_speculate_cache_rope_kernel(
                                   // head_size // 2]
     T* __restrict__ q_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens_decoder,  // [bsz]
     const float* __restrict__ cos_emb,
     const float* __restrict__ sin_emb,
@@ -355,8 +193,7 @@ __global__ void append_speculate_cache_rope_kernel(
     const int head_size,
     const int block_size,
     const int elem_cnt,
-    const int gqa_group_size,
-    const bool rope_3d) {
+    const int gqa_group_size) {
   using LoadT = AlignedVector<T, VecSize>;
   using LoadFloat = AlignedVector<float, VecSize>;
   using LoadInT = AlignedVector<InT, VecSize>;
@@ -377,12 +214,12 @@ __global__ void append_speculate_cache_rope_kernel(
        linear_index < elem_cnt;
        linear_index += step) {
     const int token_id = linear_index / hidden_size;
-    const int ori_bi = batch_id_per_token[token_id];
+    const int ori_bi = (token_id + padding_offsets[token_id]) / max_seq_len;
     if (seq_lens_decoder[ori_bi] == 0) continue;
     const int bias = linear_index % hidden_size;
     const int hi = bias / head_size;  // q + k + v
     const int h_bias = bias % head_size;
-    const int start_token_idx = cu_seqlens_q[ori_bi];
+    const int start_token_idx = ori_bi * max_seq_len - cum_offsets[ori_bi];
     const int write_seq_id =
         seq_lens_decoder[ori_bi] + token_id - start_token_idx;
     if (write_seq_id == 0) continue;
@@ -398,7 +235,7 @@ __global__ void append_speculate_cache_rope_kernel(
           ori_bi,
           seq_lens_decoder[ori_bi],
           token_id,
-          cu_seqlens_q[ori_bi]);
+          cum_offsets[ori_bi]);
     }
     const int block_offset = write_seq_id % block_size;
 
@@ -416,9 +253,8 @@ __global__ void append_speculate_cache_rope_kernel(
     if (hi < num_heads + gqa_group_size) {
       // q k rope
       const int64_t emb_idx = write_seq_id * half_head_size + h_bias / 2;
-      int64_t new_emb_idx = rope_3d ? emb_idx + ori_bi * max_seq_len * head_size : emb_idx;
-      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
     }
 #pragma unroll
     for (int i = 0; i < HalfVecSize; i++) {
@@ -475,8 +311,8 @@ __global__ void append_speculate_cache_neox_rope_kernel(
                                   // head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens_decoder,  // [bsz]
     const float* __restrict__ cos_emb,
     const float* __restrict__ sin_emb,
@@ -490,8 +326,7 @@ __global__ void append_speculate_cache_neox_rope_kernel(
     const int head_size,
     const int block_size,
     const int elem_cnt,
-    const int gqa_group_size,
-    const bool rope_3d) {
+    const int gqa_group_size) {
   using LoadT = AlignedVector<T, VecSize>;
   using LoadFloat = AlignedVector<float, VecSize>;
   using LoadInT = AlignedVector<InT, VecSize>;
@@ -512,12 +347,12 @@ __global__ void append_speculate_cache_neox_rope_kernel(
        linear_index < elem_cnt;
        linear_index += step) {
     const int token_id = linear_index / half_hidden_size;
-    const int ori_bi = batch_id_per_token[token_id];
+    const int ori_bi = (token_id + padding_offsets[token_id]) / max_seq_len;
     if (seq_lens_decoder[ori_bi] == 0) continue;
     const int bias = linear_index % half_hidden_size;
     const int hi = bias / half_head_size;  // q + k + v
     const int h_bias = bias % half_head_size;
-    const int start_token_idx = cu_seqlens_q[ori_bi];
+    const int start_token_idx = ori_bi * max_seq_len - cum_offsets[ori_bi];
     const int write_seq_id =
         seq_lens_decoder[ori_bi] + token_id - start_token_idx;
     if (write_seq_id == 0) continue;
@@ -533,7 +368,7 @@ __global__ void append_speculate_cache_neox_rope_kernel(
           ori_bi,
           seq_lens_decoder[ori_bi],
           token_id,
-          cu_seqlens_q[ori_bi]);
+          cum_offsets[ori_bi]);
     }
     const int block_offset = write_seq_id % block_size;
 
@@ -555,9 +390,8 @@ __global__ void append_speculate_cache_neox_rope_kernel(
     if (hi < num_heads + gqa_group_size) {
       // q k rope
       const int64_t emb_idx = write_seq_id * head_size + h_bias;
-      int64_t new_emb_idx = rope_3d ? emb_idx + ori_bi * max_seq_len * head_size * 2: emb_idx;
-      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
     }
 #pragma unroll
     for (int i = 0; i < VecSize; i++) {
@@ -624,8 +458,8 @@ __global__ void append_speculate_cache_int8_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -642,8 +476,7 @@ __global__ void append_speculate_cache_int8_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int gqa_group_size,
-    const bool rope_3d) {
+    const int gqa_group_size) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -651,10 +484,10 @@ __global__ void append_speculate_cache_int8_rope_kernel(
   const int wid = tid / 32;
   const int lane_id = tid % 32;
   const int token_id = blockIdx.x;
+  const int ori_token_id = token_id + padding_offsets[token_id];
+  const int bid = ori_token_id / max_seq_len;
 
-  const int bid = batch_id_per_token[token_id];
-
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - cum_offsets[bid];
   const int head_idx = blockIdx.y * NUM_WARPS + wid;
   int q_head_idx, k_head_idx, v_idx;
   const int64_t hidden_size = (num_heads + 2 * gqa_group_size) * HeadDim;
@@ -689,9 +522,8 @@ __global__ void append_speculate_cache_int8_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
       if (qkv_out_scales) {
         Load<float, VecSize>(&qkv_out_scales[bias_idx], &out_scale_vec);
       }
@@ -751,11 +583,10 @@ __global__ void append_speculate_cache_int8_rope_kernel(
     T scale;
     if (head_idx < num_heads + gqa_group_size) {
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-      Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
-      Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-      Load<float, 1>(&sin_emb[new_emb_idx + 4], &sin_emb_vec2);
+      Load<float, 1>(&cos_emb[emb_idx], &cos_emb_vec1);
+      Load<float, 1>(&cos_emb[emb_idx + 4], &cos_emb_vec2);
+      Load<float, 1>(&sin_emb[emb_idx], &sin_emb_vec1);
+      Load<float, 1>(&sin_emb[emb_idx + 4], &sin_emb_vec2);
       scale = __ldg(&cache_k_scales[kv_head_idx]);
     } else {
       scale = __ldg(&cache_v_scales[kv_head_idx]);
@@ -859,8 +690,8 @@ __global__ void append_speculate_cache_int8_neox_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -877,8 +708,7 @@ __global__ void append_speculate_cache_int8_neox_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int gqa_group_size,
-    const bool rope_3d) {
+    const int gqa_group_size) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -886,10 +716,10 @@ __global__ void append_speculate_cache_int8_neox_rope_kernel(
   const int wid = tid / 32;
   const int lane_id = tid % 32;
   const int token_id = blockIdx.x;
+  const int ori_token_id = token_id + padding_offsets[token_id];
+  const int bid = ori_token_id / max_seq_len;
 
-  const int bid = batch_id_per_token[token_id];
-
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - cum_offsets[bid];
   const int head_idx = blockIdx.y * NUM_WARPS + wid;
   int q_head_idx, k_head_idx, v_idx;
 
@@ -927,9 +757,8 @@ __global__ void append_speculate_cache_int8_neox_rope_kernel(
 
       // q rope
       const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
-      Load<float, VecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, VecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, VecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, VecSize>(&sin_emb[emb_idx], &sin_emb_vec);
       if (qkv_out_scales) {
         Load<float, VecSize>(&qkv_out_scales[bias_idx_left],
                              &left_out_scale_vec);
@@ -1024,11 +853,10 @@ __global__ void append_speculate_cache_int8_neox_rope_kernel(
 
         T scale;
         const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-        uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim * 2 : emb_idx;
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx + 8], &cos_emb_vec2);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx + 8], &sin_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx], &cos_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx + 8], &cos_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx], &sin_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx + 8], &sin_emb_vec2);
         scale = __ldg(&cache_k_scales[kv_head_idx]);
 #pragma unroll
         for (int i = 0; i < HALF_K_VEC_SIZE; i++) {
@@ -1240,8 +1068,8 @@ __global__ void append_speculate_cache_int4_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -1260,8 +1088,7 @@ __global__ void append_speculate_cache_int4_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int gqa_group_size,
-    const bool rope_3d) {
+    const int gqa_group_size) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -1270,10 +1097,10 @@ __global__ void append_speculate_cache_int4_rope_kernel(
   const int lane_id = tid % 32;
 
   const int token_id = blockIdx.x;
+  const int ori_token_id = token_id + padding_offsets[token_id];
+  const int bid = ori_token_id / max_seq_len;
 
-  const int bid = batch_id_per_token[token_id];
-
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - cum_offsets[bid];
   const int head_idx = blockIdx.y * NUM_WARPS + wid;
 
   const int64_t hidden_size = (num_heads + 2 * gqa_group_size) * HeadDim;
@@ -1303,10 +1130,6 @@ __global__ void append_speculate_cache_int4_rope_kernel(
     LoadOutScaleT out_scale_vec;
     LoadEmbT cos_emb_vec;
     LoadEmbT sin_emb_vec;
-#pragma unroll
-    for (int v_i = 0; v_i < VecSize; v_i++) {
-      bias_vec[v_i] = 0;
-    }
     const InT* qkv_now = quant_qkv + token_id * hidden_size;
     T* qkv_out_now = qkv_out + token_id * hidden_size;
 #pragma unroll
@@ -1314,22 +1137,21 @@ __global__ void append_speculate_cache_int4_rope_kernel(
          head_bias += 32 * VecSize) {
       const int bias_idx = head_idx * HeadDim + head_bias;
       Load<InT, VecSize>(&qkv_now[bias_idx], &src_vec);
-      // Load<T, VecSize>(&qkv_biases[bias_idx], &bias_vec);
-      // Load<float, VecSize>(&qkv_out_scales[bias_idx], &out_scale_vec);
+      Load<T, VecSize>(&qkv_biases[bias_idx], &bias_vec);
+      Load<float, VecSize>(&qkv_out_scales[bias_idx], &out_scale_vec);
       // q rope
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, HalfVecSize>(&cos_emb[new_emb_idx], &cos_emb_vec);
-      Load<float, HalfVecSize>(&sin_emb[new_emb_idx], &sin_emb_vec);
+      Load<float, HalfVecSize>(&cos_emb[emb_idx], &cos_emb_vec);
+      Load<float, HalfVecSize>(&sin_emb[emb_idx], &sin_emb_vec);
 #pragma unroll
       for (int i = 0; i < HalfVecSize; i++) {
         // dequant + add_bias + rope
         float input_left = static_cast<float>(src_vec[2 * i]);
         float input_right = static_cast<float>(src_vec[2 * i + 1]);
-        // input_left = input_left * out_scale_vec[2 * i] +
-        //              static_cast<float>(bias_vec[2 * i]);
-        // input_right = input_right * out_scale_vec[2 * i + 1] +
-        //               static_cast<float>(bias_vec[2 * i + 1]);
+        input_left = input_left * out_scale_vec[2 * i] +
+                     static_cast<float>(bias_vec[2 * i]);
+        input_right = input_right * out_scale_vec[2 * i + 1] +
+                      static_cast<float>(bias_vec[2 * i + 1]);
         const float cos_tmp = cos_emb_vec[i];
         const float sin_tmp = sin_emb_vec[i];
         bias_vec[2 * i] =
@@ -1345,35 +1167,6 @@ __global__ void append_speculate_cache_int4_rope_kernel(
     using LoadPadKVT = AlignedVector<uint8_t, KV_VEC_SIZE>;
     const uint32_t kv_head_idx = (head_idx - num_heads) % gqa_group_size;
 
-    if (block_offset == 0) {
-      // pad zero for this kv_head_idx for this block
-      LoadPadKVT pad_cache_vec;
-      *(reinterpret_cast<uint4*>(pad_cache_vec.val)) = make_uint4(0, 0, 0, 0);
-      if (head_idx < num_heads + gqa_group_size) {
-        constexpr int num_vecs_per_head_dim = half_head_size / KV_VEC_SIZE;
-        constexpr int num_token_each_time = 32 / num_vecs_per_head_dim;
-        const uint32_t tgt_idx = (block_idx * gqa_group_size + kv_head_idx) *
-                                     block_size * half_head_size +
-                                 lane_id % num_vecs_per_head_dim * KV_VEC_SIZE;
-        for (int block_i = lane_id / num_vecs_per_head_dim;
-             block_i < block_size;
-             block_i += num_token_each_time) {
-          Store<uint8_t, KV_VEC_SIZE>(
-              pad_cache_vec, &key_cache[tgt_idx + block_i * half_head_size]);
-        }
-      } else {
-        const int num_vecs_per_head_dim = half_block_size / KV_VEC_SIZE;
-        const int num_token_each_time = 32 / num_vecs_per_head_dim;
-        const uint32_t tgt_idx = (block_idx * gqa_group_size + kv_head_idx) *
-                                     HeadDim * half_block_size +
-                                 lane_id % num_vecs_per_head_dim * KV_VEC_SIZE;
-        for (int block_i = lane_id / num_vecs_per_head_dim; block_i < HeadDim;
-             block_i += num_token_each_time) {
-          Store<uint8_t, KV_VEC_SIZE>(
-              pad_cache_vec, &value_cache[tgt_idx + block_i * half_block_size]);
-        }
-      }
-    }
     constexpr int K_VEC_SIZE = 4;
     constexpr int HALF_K_VEC_SIZE = 2;
     using LoadKVResT = AlignedVector<uint8_t, K_VEC_SIZE>;
@@ -1389,11 +1182,7 @@ __global__ void append_speculate_cache_int4_rope_kernel(
     LoadScaleT zp_vec1, zp_vec2;
     LoadEmbT cos_emb_vec1, cos_emb_vec2;
     LoadEmbT sin_emb_vec1, sin_emb_vec2;
-#pragma unroll
-    for (int v_i = 0; v_i < HALF_K_VEC_SIZE; v_i++) {
-      bias_vec1[v_i] = 0;
-      bias_vec2[v_i] = 0;
-    }
+
     const InT* qkv_now = quant_qkv + token_id * hidden_size;
     const int head_bias = lane_id / 4 * 16 + lane_id % 4 * 2;
     //////////
@@ -1402,18 +1191,17 @@ __global__ void append_speculate_cache_int4_rope_kernel(
     Load<InT, HALF_K_VEC_SIZE>(&qkv_now[bias_idx], &src_vec1);
     Load<InT, HALF_K_VEC_SIZE>(&qkv_now[bias_idx + 8], &src_vec2);
     /////
-    // Load<T, HALF_K_VEC_SIZE>(&qkv_biases[bias_idx], &bias_vec1);
-    // Load<T, HALF_K_VEC_SIZE>(&qkv_biases[bias_idx + 8], &bias_vec2);
-    // Load<float, HALF_K_VEC_SIZE>(&qkv_out_scales[bias_idx], &out_scale_vec1);
-    // Load<float, HALF_K_VEC_SIZE>(&qkv_out_scales[bias_idx + 8],
-    //                              &out_scale_vec2);
+    Load<T, HALF_K_VEC_SIZE>(&qkv_biases[bias_idx], &bias_vec1);
+    Load<T, HALF_K_VEC_SIZE>(&qkv_biases[bias_idx + 8], &bias_vec2);
+    Load<float, HALF_K_VEC_SIZE>(&qkv_out_scales[bias_idx], &out_scale_vec1);
+    Load<float, HALF_K_VEC_SIZE>(&qkv_out_scales[bias_idx + 8],
+                                 &out_scale_vec2);
     if (head_idx < num_heads + gqa_group_size) {
       const uint32_t emb_idx = write_seq_id * half_head_size + head_bias / 2;
-      uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-      Load<float, 1>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-      Load<float, 1>(&cos_emb[new_emb_idx + 4], &cos_emb_vec2);
-      Load<float, 1>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-      Load<float, 1>(&sin_emb[new_emb_idx + 4], &sin_emb_vec2);
+      Load<float, 1>(&cos_emb[emb_idx], &cos_emb_vec1);
+      Load<float, 1>(&cos_emb[emb_idx + 4], &cos_emb_vec2);
+      Load<float, 1>(&sin_emb[emb_idx], &sin_emb_vec1);
+      Load<float, 1>(&sin_emb[emb_idx + 4], &sin_emb_vec2);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_scales[cache_idx], &scale_vec1);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_scales[cache_idx + 8], &scale_vec2);
       Load<T, HALF_K_VEC_SIZE>(&cache_k_zero_points[cache_idx], &zp_vec1);
@@ -1427,10 +1215,10 @@ __global__ void append_speculate_cache_int4_rope_kernel(
 
     float input_left = static_cast<float>(src_vec1[0]);
     float input_right = static_cast<float>(src_vec1[1]);
-    // input_left =
-    //     input_left * out_scale_vec1[0] + static_cast<float>(bias_vec1[0]);
-    // input_right =
-    //     input_right * out_scale_vec1[1] + static_cast<float>(bias_vec1[1]);
+    input_left =
+        input_left * out_scale_vec1[0] + static_cast<float>(bias_vec1[0]);
+    input_right =
+        input_right * out_scale_vec1[1] + static_cast<float>(bias_vec1[1]);
     if (head_idx < num_heads + gqa_group_size) {
       float cos_tmp = cos_emb_vec1[0];
       float sin_tmp = sin_emb_vec1[0];
@@ -1445,10 +1233,10 @@ __global__ void append_speculate_cache_int4_rope_kernel(
 
     input_left = static_cast<float>(src_vec2[0]);
     input_right = static_cast<float>(src_vec2[1]);
-    // input_left =
-    //     input_left * out_scale_vec2[0] + static_cast<float>(bias_vec2[0]);
-    // input_right =
-    //     input_right * out_scale_vec2[1] + static_cast<float>(bias_vec2[1]);
+    input_left =
+        input_left * out_scale_vec2[0] + static_cast<float>(bias_vec2[0]);
+    input_right =
+        input_right * out_scale_vec2[1] + static_cast<float>(bias_vec2[1]);
     if (head_idx < num_heads + gqa_group_size) {
       float cos_tmp = cos_emb_vec2[0];
       float sin_tmp = sin_emb_vec2[0];
@@ -1586,8 +1374,8 @@ __global__ void append_speculate_cache_int4_neox_rope_kernel(
                                         // block_size, head_size // 2]
     T* __restrict__ qkv_out,
     const int* __restrict__ block_tables,     // [bsz, max_blocks_per_seq]
-    const int* __restrict__ batch_id_per_token,  // [num_tokens]
-    const int* __restrict__ cu_seqlens_q,
+    const int* __restrict__ padding_offsets,  // [num_tokens]
+    const int* __restrict__ cum_offsets,
     const int* __restrict__ seq_lens,          // [bsz]
     const int* __restrict__ seq_lens_encoder,  // [bsz]
     const float* __restrict__ cos_emb,
@@ -1606,8 +1394,7 @@ __global__ void append_speculate_cache_int4_neox_rope_kernel(
     const int block_size,
     const float max_bound,
     const float min_bound,
-    const int gqa_group_size,
-    const bool rope_3d) {
+    const int gqa_group_size) {
   static_assert(HeadDim == 128, "just support HeadDim be 128 now!");
   static_assert(VecSize == 4, "just support VecSize be 4 now, 32 * 4!");
   constexpr int NUM_WARPS = 4;
@@ -1616,10 +1403,10 @@ __global__ void append_speculate_cache_int4_neox_rope_kernel(
   const int lane_id = tid % 32;
 
   const int token_id = blockIdx.x;
+  const int ori_token_id = token_id + padding_offsets[token_id];
+  const int bid = ori_token_id / max_seq_len;
 
-  const int bid = batch_id_per_token[token_id];
-
-  const int start_token_idx = cu_seqlens_q[bid];
+  const int start_token_idx = bid * max_seq_len - cum_offsets[bid];
   const int head_idx = blockIdx.y * NUM_WARPS + wid;
 
   const int64_t hidden_size = (num_heads + 2 * gqa_group_size) * HeadDim;
@@ -1757,11 +1544,10 @@ __global__ void append_speculate_cache_int4_neox_rope_kernel(
                                      &right_out_scale_vec2);
 
         const uint32_t emb_idx = write_seq_id * HeadDim + head_bias;
-        uint32_t new_emb_idx = rope_3d ? emb_idx + bid * max_seq_len * HeadDim : emb_idx;
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx], &cos_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&cos_emb[new_emb_idx + 8], &cos_emb_vec2);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx], &sin_emb_vec1);
-        Load<float, HALF_K_VEC_SIZE>(&sin_emb[new_emb_idx + 8], &sin_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx], &cos_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&cos_emb[emb_idx + 8], &cos_emb_vec2);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx], &sin_emb_vec1);
+        Load<float, HALF_K_VEC_SIZE>(&sin_emb[emb_idx + 8], &sin_emb_vec2);
         Load<T, HALF_K_VEC_SIZE>(&cache_k_scales[left_cache_idx],
                                  &left_scale_vec1);
         Load<T, HALF_K_VEC_SIZE>(&cache_k_scales[left_cache_idx + 8],
@@ -2006,4 +1792,4 @@ __global__ void append_speculate_cache_int4_neox_rope_kernel(
           (uint_quant_value2 << 4) | (uint_quant_value1 & 0x0F);
     }
   }
-}
+}

custom_ops/gpu_ops/append_attn/speculate_write_cache_with_rope_kernel.cu

@@ -15,78 +15,6 @@
 #include "speculate_write_cache_with_rope_kernel.h"
 #include "utils.cuh"
 
-template <typename T, typename QKV_TYPE>
-void append_speculate_cache_rope_qk_norm(const QKV_TYPE* qkv,
-                              T* key_cache,
-                              T* value_cache,
-                              T* qkv_out,
-                              const int* block_tables,
-                              const int* batch_id_per_token,
-                              const int* cu_seqlens_q,
-                              const int* seq_lens,
-                              const int* seq_lens_encoder,
-                              const float* cos_emb,
-                              const float* sin_emb,
-                              const float* qkv_out_scales,
-                              const T* qkv_biases,
-                              const int max_seq_len,
-                              const int max_blocks_per_seq,
-                              const int num_heads,
-                              const int kv_num_heads,
-                              const int dim_head,
-                              const int block_size,
-                              const int bsz,
-                              const int token_num,
-                              const cudaStream_t& stream,
-                              const bool use_neox_style,
-                              const float* q_norm_weight,
-                              const float* k_norm_weight,
-                              const float rms_norm_eps,
-                              const bool rope_3d) {
-  int output_inner_dim = num_heads + 2 * kv_num_heads;
-  const uint32_t elem_nums =
-      use_neox_style ? token_num * (num_heads + 2 * kv_num_heads) * dim_head / 2
-                     : token_num * (num_heads + 2 * kv_num_heads) * dim_head;
-  constexpr int HEAD_DIM = 128;
-
-  constexpr int PackSize = HEAD_DIM / kWarpSize;
-  const int pack_num = elem_nums / PackSize;
-  const int blocksize = 128;
-  int grid_size = 1;
-  GetNumBlocks<128>(pack_num, &grid_size);
-  if (use_neox_style) {
-      PD_THROW(
-          "append_speculate_cache_rope_qk_norm not support neox rope yet");
-  } else {
-    dim3 block_dim(kWarpSize, blocksize / kWarpSize, 1);
-    append_speculate_cache_T_rope_qk_norm_kernel<T, PackSize>
-      <<<grid_size, block_dim, 0, stream>>>(qkv,
-                                            key_cache,
-                                            value_cache,
-                                            qkv_out,
-                                            block_tables,
-                                            batch_id_per_token,
-                                            cu_seqlens_q,
-                                            seq_lens,
-                                            cos_emb,
-                                            sin_emb,
-                                            qkv_out_scales,
-                                            qkv_biases,
-                                            max_seq_len,
-                                            max_blocks_per_seq,
-                                            num_heads,
-                                            output_inner_dim,
-                                            dim_head,
-                                            block_size,
-                                            elem_nums,
-                                            kv_num_heads,
-                                            q_norm_weight,
-                                            k_norm_weight,
-                                            rms_norm_eps,
-                                            rope_3d);
-  }
-}
-
 // rope + write
 template <typename T, typename QKV_TYPE>
 void append_speculate_cache_rope(const QKV_TYPE* qkv,
@@ -94,8 +22,8 @@ void append_speculate_cache_rope(const QKV_TYPE* qkv,
                                  T* value_cache,
                                  T* qkv_out,
                                  const int* block_tables,
-                                 const int* batch_id_per_token,
-                                 const int* cu_seqlens_q,
+                                 const int* padding_offsets,
+                                 const int* cum_offsets,
                                  const int* seq_lens,
                                  const int* seq_lens_encoder,
                                  const float* cos_emb,
@@ -111,8 +39,7 @@ void append_speculate_cache_rope(const QKV_TYPE* qkv,
                                  const int bsz,
                                  const int token_num,
                                  const cudaStream_t& stream,
-                                 const bool use_neox_style,
-                                 const bool rope_3d) {
+                                 const bool use_neox_style) {
   int output_inner_dim = num_heads + 2 * kv_num_heads;
 
   const uint32_t elem_nums =
@@ -132,8 +59,8 @@ void append_speculate_cache_rope(const QKV_TYPE* qkv,
             value_cache,
             qkv_out,
             block_tables,
-            batch_id_per_token,
-            cu_seqlens_q,
+            padding_offsets,
+            cum_offsets,
             seq_lens,
             cos_emb,
             sin_emb,
@@ -146,8 +73,7 @@ void append_speculate_cache_rope(const QKV_TYPE* qkv,
             dim_head,
             block_size,
             elem_nums,
-            kv_num_heads,
-            rope_3d);
+            kv_num_heads);
   } else {
     append_speculate_cache_rope_kernel<T, PackSize>
         <<<grid_size, threads_per_block, 0, stream>>>(
@@ -156,8 +82,8 @@ void append_speculate_cache_rope(const QKV_TYPE* qkv,
             value_cache,
             qkv_out,
             block_tables,
-            batch_id_per_token,
-            cu_seqlens_q,
+            padding_offsets,
+            cum_offsets,
             seq_lens,
             cos_emb,
             sin_emb,
@@ -170,8 +96,7 @@ void append_speculate_cache_rope(const QKV_TYPE* qkv,
             dim_head,
             block_size,
             elem_nums,
-            kv_num_heads,
-            rope_3d);
+            kv_num_heads);
   }
 }
 
@@ -181,8 +106,8 @@ void append_speculate_cache_int8_rope(const QKV_TYPE* qkv,
                                       uint8_t* value_cache,
                                       T* qkv_out,
                                       const int* block_tables,
-                                      const int* batch_id_per_token,
-                                      const int* cu_seqlens_q,
+                                      const int* padding_offsets,
+                                      const int* cum_offsets,
                                       const int* seq_lens,
                                       const int* seq_lens_encoder,
                                       const float* cos_emb,
@@ -200,8 +125,7 @@ void append_speculate_cache_int8_rope(const QKV_TYPE* qkv,
                                       const int bsz,
                                       const int token_num,
                                       const cudaStream_t& stream,
-                                      const bool use_neox_style,
-                                      const bool rope_3d) {
+                                      const bool use_neox_style) {
   constexpr int num_warps = 4;
   const int all_warps =
       ((num_heads + 2 * kv_num_heads) + num_warps - 1) / num_warps * num_warps;
@@ -212,8 +136,8 @@ void append_speculate_cache_int8_rope(const QKV_TYPE* qkv,
                                              value_cache,
                                              seq_lens,
                                              block_tables,
-                                             batch_id_per_token,
-                                             cu_seqlens_q,
+                                             padding_offsets,
+                                             cum_offsets,
                                              seq_lens_encoder,
                                              max_seq_len,
                                              max_blocks_per_seq,
@@ -227,8 +151,8 @@ void append_speculate_cache_int8_rope(const QKV_TYPE* qkv,
                                                value_cache,
                                                qkv_out,
                                                block_tables,
-                                               batch_id_per_token,
-                                               cu_seqlens_q,
+                                               padding_offsets,
+                                               cum_offsets,
                                                seq_lens,
                                                seq_lens_encoder,
                                                cos_emb,
@@ -243,8 +167,7 @@ void append_speculate_cache_int8_rope(const QKV_TYPE* qkv,
                                                block_size,
                                                127.0f,
                                                -127.0f,
-                                               kv_num_heads,
-                                               rope_3d);
+                                               kv_num_heads);
   } else {
     append_speculate_cache_int8_rope_kernel<T, 4, 0, 128, QKV_TYPE, IsFP8>
         <<<grids, num_warps * 32, 0, stream>>>(qkv,
@@ -252,8 +175,8 @@ void append_speculate_cache_int8_rope(const QKV_TYPE* qkv,
                                                value_cache,
                                                qkv_out,
                                                block_tables,
-                                               batch_id_per_token,
-                                               cu_seqlens_q,
+                                               padding_offsets,
+                                               cum_offsets,
                                                seq_lens,
                                                seq_lens_encoder,
                                                cos_emb,
@@ -268,8 +191,7 @@ void append_speculate_cache_int8_rope(const QKV_TYPE* qkv,
                                                block_size,
                                                127.0f,
                                                -127.0f,
-                                               kv_num_heads,
-                                               rope_3d);
+                                               kv_num_heads);
   }
 }
 
@@ -279,8 +201,8 @@ void append_speculate_cache_int4_rope(const QKV_TYPE* qkv,
                                       uint8_t* value_cache,
                                       T* qkv_out,
                                       const int* block_tables,
-                                      const int* batch_id_per_token,
-                                      const int* cu_seqlens_q,
+                                      const int* padding_offsets,
+                                      const int* cum_offsets,
                                       const int* seq_lens,
                                       const int* seq_lens_encoder,
                                       const float* cos_emb,
@@ -300,8 +222,7 @@ void append_speculate_cache_int4_rope(const QKV_TYPE* qkv,
                                       const int bsz,
                                       const int token_num,
                                       const cudaStream_t& stream,
-                                      const bool use_neox_style,
-                                      const bool rope_3d) {
+                                      const bool use_neox_style) {
   constexpr int num_warps = 4;
   const int all_warps =
       ((num_heads + 2 * kv_num_heads) + num_warps - 1) / num_warps * num_warps;
@@ -312,8 +233,8 @@ void append_speculate_cache_int4_rope(const QKV_TYPE* qkv,
                                              value_cache,
                                              seq_lens,
                                              block_tables,
-                                             batch_id_per_token,
-                                             cu_seqlens_q,
+                                             padding_offsets,
+                                             cum_offsets,
                                              seq_lens_encoder,
                                              max_seq_len,
                                              max_blocks_per_seq,
@@ -327,8 +248,8 @@ void append_speculate_cache_int4_rope(const QKV_TYPE* qkv,
                                                value_cache,
                                                qkv_out,
                                                block_tables,
-                                               batch_id_per_token,
-                                               cu_seqlens_q,
+                                               padding_offsets,
+                                               cum_offsets,
                                                seq_lens,
                                                seq_lens_encoder,
                                                cos_emb,
@@ -345,8 +266,7 @@ void append_speculate_cache_int4_rope(const QKV_TYPE* qkv,
                                                block_size,
                                                7.0f,
                                                -8.0f,
-                                               kv_num_heads,
-                                               rope_3d);
+                                               kv_num_heads);
   } else {
     append_speculate_cache_int4_rope_kernel<T, 4>
         <<<grids, num_warps * 32, 0, stream>>>(qkv,
@@ -354,8 +274,8 @@ void append_speculate_cache_int4_rope(const QKV_TYPE* qkv,
                                                value_cache,
                                                qkv_out,
                                                block_tables,
-                                               batch_id_per_token,
-                                               cu_seqlens_q,
+                                               padding_offsets,
+                                               cum_offsets,
                                                seq_lens,
                                                seq_lens_encoder,
                                                cos_emb,
@@ -372,8 +292,7 @@ void append_speculate_cache_int4_rope(const QKV_TYPE* qkv,
                                                block_size,
                                                7.0f,
                                                -8.0f,
-                                               kv_num_heads,
-                                               rope_3d);
+                                               kv_num_heads);
   }
 }
 template <typename T, typename QKV_TYPE>
@@ -382,8 +301,8 @@ void SpeculateWriteCacheWithRoPEKernel(
     const paddle::Tensor& qkv,
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& qkv_out_scales,
@@ -394,15 +313,11 @@ void SpeculateWriteCacheWithRoPEKernel(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
-    const bool rope_3d,
     const int max_seq_len,
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps) {
+    paddle::Tensor* value_cache_out) {
   typedef cascade_attn_type_traits<T> traits_;
   typedef cascade_attn_type_traits<QKV_TYPE> qkt_nv_type_;
   typedef typename traits_::type DataType_;
@@ -427,185 +342,142 @@ void SpeculateWriteCacheWithRoPEKernel(
             ? rotary_embs.get().data<float>() + max_seq_len * dim_head
             : rotary_embs.get().data<float>() + max_seq_len * dim_head / 2;
   }
-  if (q_norm_weight && k_norm_weight) {
-    if (cache_quant_type_str == "none") {
-      append_speculate_cache_rope_qk_norm(
-          reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-          reinterpret_cast<DataType_*>(key_cache_out->data<T>()),
-          reinterpret_cast<DataType_*>(value_cache_out->data<T>()),
-          reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-          block_tables.data<int>(),
-          batch_id_per_token.data<int>(),
-          cu_seqlens_q.data<int>(),
-          seq_lens.data<int>(),
-          seq_lens_encoder.data<int>(),
-          cos_emb,
-          sin_emb,
-          qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-          qkv_biases ? reinterpret_cast<DataType_*>(
-                          const_cast<T*>(qkv_biases.get().data<T>()))
-                    : nullptr,
-          max_seq_len,
-          max_blocks_per_seq,
-          num_heads,
-          kv_num_heads,
-          dim_head,
-          block_size,
-          bsz,
-          token_nums,
-          stream,
-          use_neox_rotary_style,
-          reinterpret_cast<const float*>(q_norm_weight.get().data<float>()),
-          reinterpret_cast<const float*>(k_norm_weight.get().data<float>()),
-          rms_norm_eps,
-          rope_3d);
-    } else {
-      PD_THROW(
-          "append_decode_cache_rope_qk_norm not support cachekv quant yet");
-    }
-
+  if (cache_quant_type_str == "none") {
+    append_speculate_cache_rope(
+        reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+        reinterpret_cast<DataType_*>(key_cache_out->data<T>()),
+        reinterpret_cast<DataType_*>(value_cache_out->data<T>()),
+        reinterpret_cast<DataType_*>(qkv_out->data<T>()),
+        block_tables.data<int>(),
+        padding_offsets.data<int>(),
+        cum_offsets.data<int>(),
+        seq_lens.data<int>(),
+        seq_lens_encoder.data<int>(),
+        cos_emb,
+        sin_emb,
+        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+        qkv_biases ? reinterpret_cast<DataType_*>(
+                         const_cast<T*>(qkv_biases.get().data<T>()))
+                   : nullptr,
+        max_seq_len,
+        max_blocks_per_seq,
+        num_heads,
+        kv_num_heads,
+        dim_head,
+        block_size,
+        bsz,
+        token_nums,
+        stream,
+        use_neox_rotary_style);
+  } else if (cache_quant_type_str == "cache_int8") {
+    append_speculate_cache_int8_rope(
+        reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+        key_cache_out->data<uint8_t>(),
+        value_cache_out->data<uint8_t>(),
+        reinterpret_cast<DataType_*>(qkv_out->data<T>()),
+        block_tables.data<int>(),
+        padding_offsets.data<int>(),
+        cum_offsets.data<int>(),
+        seq_lens.data<int>(),
+        seq_lens_encoder.data<int>(),
+        cos_emb,
+        sin_emb,
+        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+        qkv_biases ? reinterpret_cast<DataType_*>(
+                         const_cast<T*>(qkv_biases.get().data<T>()))
+                   : nullptr,
+        cache_k_scale ? reinterpret_cast<DataType_*>(
+                            const_cast<T*>(cache_k_scale.get().data<T>()))
+                      : nullptr,
+        cache_v_scale ? reinterpret_cast<DataType_*>(
+                            const_cast<T*>(cache_v_scale.get().data<T>()))
+                      : nullptr,
+        max_seq_len,
+        max_blocks_per_seq,
+        num_heads,
+        kv_num_heads,
+        dim_head,
+        block_size,
+        bsz,
+        token_nums,
+        stream,
+        use_neox_rotary_style);
+  } else if (cache_quant_type_str == "cache_fp8") {
+    append_speculate_cache_int8_rope<DataType_, QKV_TYPE, true>(
+        reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+        key_cache_out->data<uint8_t>(),
+        value_cache_out->data<uint8_t>(),
+        reinterpret_cast<DataType_*>(qkv_out->data<T>()),
+        block_tables.data<int>(),
+        padding_offsets.data<int>(),
+        cum_offsets.data<int>(),
+        seq_lens.data<int>(),
+        seq_lens_encoder.data<int>(),
+        cos_emb,
+        sin_emb,
+        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+        qkv_biases ? reinterpret_cast<DataType_*>(
+                         const_cast<T*>(qkv_biases.get().data<T>()))
+                   : nullptr,
+        cache_k_scale ? reinterpret_cast<DataType_*>(
+                            const_cast<T*>(cache_k_scale.get().data<T>()))
+                      : nullptr,
+        cache_v_scale ? reinterpret_cast<DataType_*>(
+                            const_cast<T*>(cache_v_scale.get().data<T>()))
+                      : nullptr,
+        max_seq_len,
+        max_blocks_per_seq,
+        num_heads,
+        kv_num_heads,
+        dim_head,
+        block_size,
+        bsz,
+        token_nums,
+        stream,
+        use_neox_rotary_style);
+  } else if (cache_quant_type_str == "cache_int4_zp") {
+    append_speculate_cache_int4_rope(
+        reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
+        key_cache_out->data<uint8_t>(),
+        value_cache_out->data<uint8_t>(),
+        reinterpret_cast<DataType_*>(const_cast<T*>(qkv_out->data<T>())),
+        block_tables.data<int>(),
+        padding_offsets.data<int>(),
+        cum_offsets.data<int>(),
+        seq_lens.data<int>(),
+        seq_lens_encoder.data<int>(),
+        cos_emb,
+        sin_emb,
+        qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
+        qkv_biases ? reinterpret_cast<DataType_*>(
+                         const_cast<T*>(qkv_biases.get().data<T>()))
+                   : nullptr,
+        cache_k_scale ? reinterpret_cast<DataType_*>(
+                            const_cast<T*>(cache_k_scale.get().data<T>()))
+                      : nullptr,
+        cache_v_scale ? reinterpret_cast<DataType_*>(
+                            const_cast<T*>(cache_v_scale.get().data<T>()))
+                      : nullptr,
+        cache_k_zp ? reinterpret_cast<DataType_*>(
+                         const_cast<T*>(cache_k_zp.get().data<T>()))
+                   : nullptr,
+        cache_v_zp ? reinterpret_cast<DataType_*>(
+                         const_cast<T*>(cache_v_zp.get().data<T>()))
+                   : nullptr,
+        max_seq_len,
+        max_blocks_per_seq,
+        num_heads,
+        kv_num_heads,
+        dim_head,
+        block_size,
+        bsz,
+        token_nums,
+        stream,
+        use_neox_rotary_style);
   } else {
-    if (cache_quant_type_str == "none") {
-        append_speculate_cache_rope(
-            reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-            reinterpret_cast<DataType_*>(key_cache_out->data<T>()),
-            reinterpret_cast<DataType_*>(value_cache_out->data<T>()),
-            reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-            block_tables.data<int>(),
-            batch_id_per_token.data<int>(),
-            cu_seqlens_q.data<int>(),
-            seq_lens.data<int>(),
-            seq_lens_encoder.data<int>(),
-            cos_emb,
-            sin_emb,
-            qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-            qkv_biases ? reinterpret_cast<DataType_*>(
-                            const_cast<T*>(qkv_biases.get().data<T>()))
-                    : nullptr,
-            max_seq_len,
-            max_blocks_per_seq,
-            num_heads,
-            kv_num_heads,
-            dim_head,
-            block_size,
-            bsz,
-            token_nums,
-            stream,
-            use_neox_rotary_style,
-            rope_3d);
-    } else if (cache_quant_type_str == "cache_int8") {
-        append_speculate_cache_int8_rope(
-            reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-            key_cache_out->data<uint8_t>(),
-            value_cache_out->data<uint8_t>(),
-            reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-            block_tables.data<int>(),
-            batch_id_per_token.data<int>(),
-            cu_seqlens_q.data<int>(),
-            seq_lens.data<int>(),
-            seq_lens_encoder.data<int>(),
-            cos_emb,
-            sin_emb,
-            qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-            qkv_biases ? reinterpret_cast<DataType_*>(
-                            const_cast<T*>(qkv_biases.get().data<T>()))
-                    : nullptr,
-            cache_k_scale ? reinterpret_cast<DataType_*>(
-                                const_cast<T*>(cache_k_scale.get().data<T>()))
-                        : nullptr,
-            cache_v_scale ? reinterpret_cast<DataType_*>(
-                                const_cast<T*>(cache_v_scale.get().data<T>()))
-                        : nullptr,
-            max_seq_len,
-            max_blocks_per_seq,
-            num_heads,
-            kv_num_heads,
-            dim_head,
-            block_size,
-            bsz,
-            token_nums,
-            stream,
-            use_neox_rotary_style,
-            rope_3d);
-    } else if (cache_quant_type_str == "cache_fp8") {
-        append_speculate_cache_int8_rope<DataType_, QKV_TYPE, true>(
-            reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-            key_cache_out->data<uint8_t>(),
-            value_cache_out->data<uint8_t>(),
-            reinterpret_cast<DataType_*>(qkv_out->data<T>()),
-            block_tables.data<int>(),
-            batch_id_per_token.data<int>(),
-            cu_seqlens_q.data<int>(),
-            seq_lens.data<int>(),
-            seq_lens_encoder.data<int>(),
-            cos_emb,
-            sin_emb,
-            qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-            qkv_biases ? reinterpret_cast<DataType_*>(
-                            const_cast<T*>(qkv_biases.get().data<T>()))
-                    : nullptr,
-            cache_k_scale ? reinterpret_cast<DataType_*>(
-                                const_cast<T*>(cache_k_scale.get().data<T>()))
-                        : nullptr,
-            cache_v_scale ? reinterpret_cast<DataType_*>(
-                                const_cast<T*>(cache_v_scale.get().data<T>()))
-                        : nullptr,
-            max_seq_len,
-            max_blocks_per_seq,
-            num_heads,
-            kv_num_heads,
-            dim_head,
-            block_size,
-            bsz,
-            token_nums,
-            stream,
-            use_neox_rotary_style,
-            rope_3d);
-    } else if (cache_quant_type_str == "cache_int4_zp") {
-        append_speculate_cache_int4_rope(
-            reinterpret_cast<const QKV_TYPE*>(qkv_ptr),
-            key_cache_out->data<uint8_t>(),
-            value_cache_out->data<uint8_t>(),
-            reinterpret_cast<DataType_*>(const_cast<T*>(qkv_out->data<T>())),
-            block_tables.data<int>(),
-            batch_id_per_token.data<int>(),
-            cu_seqlens_q.data<int>(),
-            seq_lens.data<int>(),
-            seq_lens_encoder.data<int>(),
-            cos_emb,
-            sin_emb,
-            qkv_out_scales ? qkv_out_scales.get().data<float>() : nullptr,
-            qkv_biases ? reinterpret_cast<DataType_*>(
-                            const_cast<T*>(qkv_biases.get().data<T>()))
-                    : nullptr,
-            cache_k_scale ? reinterpret_cast<DataType_*>(
-                                const_cast<T*>(cache_k_scale.get().data<T>()))
-                        : nullptr,
-            cache_v_scale ? reinterpret_cast<DataType_*>(
-                                const_cast<T*>(cache_v_scale.get().data<T>()))
-                        : nullptr,
-            cache_k_zp ? reinterpret_cast<DataType_*>(
-                            const_cast<T*>(cache_k_zp.get().data<T>()))
-                    : nullptr,
-            cache_v_zp ? reinterpret_cast<DataType_*>(
-                            const_cast<T*>(cache_v_zp.get().data<T>()))
-                    : nullptr,
-            max_seq_len,
-            max_blocks_per_seq,
-            num_heads,
-            kv_num_heads,
-            dim_head,
-            block_size,
-            bsz,
-            token_nums,
-            stream,
-            use_neox_rotary_style,
-            rope_3d);
-    } else {
-        PD_THROW(
-            "cache_quant_type_str should be one of [none, cache_int8, "
-            "cache_int4_zp]");
-    }
+    PD_THROW(
+        "cache_quant_type_str should be one of [none, cache_int8, "
+        "cache_int4_zp]");
   }
 }
 
@@ -616,8 +488,8 @@ template void SpeculateWriteCacheWithRoPEKernel<paddle::bfloat16, int>(
               // gqa_group_size, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& qkv_out_scales,
@@ -628,15 +500,11 @@ template void SpeculateWriteCacheWithRoPEKernel<paddle::bfloat16, int>(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
-    const bool rope_3d,
     const int max_seq_len,
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps);
+    paddle::Tensor* value_cache_out);
 
 template void
 SpeculateWriteCacheWithRoPEKernel<paddle::bfloat16, paddle::bfloat16>(
@@ -646,8 +514,8 @@ SpeculateWriteCacheWithRoPEKernel<paddle::bfloat16, paddle::bfloat16>(
               // gqa_group_size, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& qkv_out_scales,
@@ -658,15 +526,11 @@ SpeculateWriteCacheWithRoPEKernel<paddle::bfloat16, paddle::bfloat16>(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
-    const bool rope_3d,
     const int max_seq_len,
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps);
+    paddle::Tensor* value_cache_out);
 
 template void SpeculateWriteCacheWithRoPEKernel<paddle::float16, int>(
     const AppendAttnMetaData& meta_data,
@@ -675,8 +539,8 @@ template void SpeculateWriteCacheWithRoPEKernel<paddle::float16, int>(
               // gqa_group_size, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& qkv_out_scales,
@@ -687,15 +551,11 @@ template void SpeculateWriteCacheWithRoPEKernel<paddle::float16, int>(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
-    const bool rope_3d,
     const int max_seq_len,
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps);
+    paddle::Tensor* value_cache_out);
 
 
 template void
@@ -706,8 +566,8 @@ SpeculateWriteCacheWithRoPEKernel<paddle::float16, paddle::float16>(
               // gqa_group_size, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& qkv_out_scales,
@@ -718,12 +578,8 @@ SpeculateWriteCacheWithRoPEKernel<paddle::float16, paddle::float16>(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
-    const bool rope_3d,
     const int max_seq_len,
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps);
+    paddle::Tensor* value_cache_out);

custom_ops/gpu_ops/append_attn/speculate_write_cache_with_rope_kernel.h

@@ -23,8 +23,8 @@ void SpeculateWriteCacheWithRoPEKernel(
               // gqa_group_size, head_dim] if GQA)
     const paddle::Tensor& seq_lens,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_tables,
     const paddle::optional<paddle::Tensor>& rotary_embs,
     const paddle::optional<paddle::Tensor>& qkv_out_scales,
@@ -35,12 +35,8 @@ void SpeculateWriteCacheWithRoPEKernel(
     const paddle::optional<paddle::Tensor>& cache_v_zp,
     const std::string& cache_quant_type_str,
     const bool use_neox_rotary_style,
-    const bool rope_3d,
     const int max_seq_len,
     cudaStream_t& stream,
     paddle::Tensor* qkv_out,
     paddle::Tensor* key_cache_out,
-    paddle::Tensor* value_cache_out,
-    const paddle::optional<paddle::Tensor>& q_norm_weight,
-    const paddle::optional<paddle::Tensor>& k_norm_weight,
-    const float rms_norm_eps);
+    paddle::Tensor* value_cache_out);

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c16_bfloat16_bfloat16_kernel.cu

@@ -37,8 +37,8 @@ template void CascadeAppendAttentionC16Kernel<paddle::bfloat16, paddle::bfloat16
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c16_bfloat16_fp8_kernel.cu

@@ -36,8 +36,8 @@ template void CascadeAppendAttentionC16Kernel<paddle::bfloat16, paddle::float8_e
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c16_bfloat16_int8_kernel.cu

@@ -36,8 +36,8 @@ template void CascadeAppendAttentionC16Kernel<paddle::bfloat16, int8_t>(
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c16_float16_float16_kernel.cu

@@ -36,8 +36,8 @@ template void CascadeAppendAttentionC16Kernel<paddle::float16, paddle::float16>(
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c16_float16_fp8_kernel.cu

@@ -36,8 +36,8 @@ template void CascadeAppendAttentionC16Kernel<paddle::float16, paddle::float8_e4
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c16_float16_int8_kernel.cu

@@ -36,8 +36,8 @@ template void CascadeAppendAttentionC16Kernel<paddle::float16, int8_t>(
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c4_bfloat16_bfloat16_kernel.cu

@@ -36,8 +36,8 @@ template void CascadeAppendAttentionC4Kernel<paddle::bfloat16, paddle::bfloat16>
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,

custom_ops/gpu_ops/append_attn/template_instantiation/append_attention_c4_bfloat16_fp8_kernel.cu

@@ -36,8 +36,8 @@ template void CascadeAppendAttentionC4Kernel<paddle::bfloat16, paddle::float8_e4
     const paddle::Tensor& seq_lens_q,
     const paddle::Tensor& seq_lens_kv,
     const paddle::Tensor& seq_lens_encoder,
-    const paddle::Tensor& batch_id_per_token,
-    const paddle::Tensor& cu_seqlens_q,
+    const paddle::Tensor& padding_offsets,
+    const paddle::Tensor& cum_offsets,
     const paddle::Tensor& block_table,
     const paddle::Tensor& batch_ids,
     const paddle::Tensor& tile_ids_per_batch,